Abstract

The coupled characterization of overburden movement, surrounding-rock failure evolution, and mine-pressure response remains difficult during high-intensity longwall mining because these processes are commonly measured and interpreted using separate data streams. In this study, a multi-source sensing and interpretation framework was established for the S8310 longwall face of Yangmei No. 1 Mine by integrating physical similarity simulation, underground monitoring, UDEC numerical modeling, region-of-interest (ROI) image-feature extraction, and sliding-window long short-term memory (LSTM) analysis. Fiber Bragg grating (FBG) sensors and conventional monitoring were used to obtain key-stratum deformation and support-pressure responses. A joint-state-based damage index (DI) was constructed from fixed-ROI UDEC outputs to quantify progressive structural activation beneath the key stratum. The results indicate that the overburden evolved from global bending and local crack initiation to fracture expansion, interface degradation, and interlayer slip. In the physical model, the initial weighting interval was approximately 37.5 cm, and the periodic weighting intervals were mainly 13.3–15.7 cm; these correspond to prototype-scale distances of approximately 37.5 m and 13.3–15.7 m, respectively. Peak abutment pressure occurred approximately 9–17 cm ahead of the face at the model scale, with a peak coefficient of 1.40–1.68. When the prototype advance distance increased from 37.5 m to 80.3 m, the ROI exhibited enhanced joint activation and a transition from local slip to continuous tensile opening and slip. For the pressure-sequence analysis, the 24-step-window LSTM produced lower errors than the 36- and 48-step-window LSTM models under the same preprocessing and training settings. Additional persistence, moving-average, and random forest baselines were added to clarify the prediction context: the persistence baseline performed strongly because of the high short-term autocorrelation of the pressure series, whereas the 24-step-window LSTM outperformed the moving-average and random forest baselines. Accordingly, the LSTM module is interpreted as an auxiliary temporal-pattern identification tool rather than as an exclusive optimal predictor. The comparison between DI evolution and pressure-prediction behavior suggests that rapid DI growth is mechanically consistent with stronger pressure-sequence non-stationarity and increased prediction deviation near active mine-pressure events. The proposed framework provides a sensor-oriented and physically interpretable approach for linking overburden failure evolution with mine-pressure response in longwall mining.

IPC Classification

Keywords