Answers are grounded in this paper and cite their sources.
Reliable roof-water-inrush risk zoning in coal mines remains difficult under small-sample conditions because geological, hydrogeological, and mining-induced factors interact nonlinearly, and field labels are often derived from engineering evidence rather than complete accident records. This study aims to develop and preliminarily validate a deployment-oriented workflow for case-specific roof-water-inrush risk-zone classification in the No. 4−3 coal seam of Shiyangou Coal Mine, Shaanxi Province, China. A total of 100 spatial evaluation samples were compiled using eight indicators grouped into water-resisting capacity, water-supply intensity, water-conducting pathways, mining-induced disturbance, and rock-mass integrity. The Safe-Zone and Risk-Zone labels were generated from mine-water-prevention documents, borehole and hydrogeological evidence, abnormal goaf-water information, and expert engineering judgment. A lightweight CNN-LSTM model was trained with leakage-controlled data splitting, early stopping, L2 regularization, gradient clipping, and model-capacity control. Because the inputs are static or quasi-static spatial indicators, LSTM was used only as a gated feature-dependency module for CNN-derived feature sequences, and its necessity was examined using ablation and feature-order perturbation tests. In repeated stratified five-fold cross-validation, the model achieved mean accuracy, F1-score, and AUC values of 0.792 ± 0.108, 0.633 ± 0.111, and 0.829 ± 0.097, respectively. On an independent prediction dataset, the accuracy and AUC were 0.750 and 0.743, respectively. After ONNX export, STM32CubeMX evaluation indicated 68,962 B Flash and 3252 B RAM consumption. These results indicate the preliminary feasibility of embedded roof-water-inrush risk-zone classification for this specific mine case. However, the findings do not demonstrate cross-mine transferability; further validation using external datasets, independent labels, and field deployment records is required.
No takes yet. Share an insight, caveat, or question.
Yang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: