Information leakage, spatial generalisation and uncertainty in data-driven in situ stress assessment for coal mines
Year: 2026
Type: article
Abstract: Accurate in situ stress characterisation supports underground excavation assessment, but data-driven estimates can be biased by target-derived information and spatially unrepresentative validation. We evaluated a physics-consistent workflow separating blind prediction, conditional stress completion and consistency screening using 725 hydraulic-fracturing records assigned to 33 coalfield-scale database groups and 36 representative positions. Complete coalfield and coordinate groups were held out, with repeated validation, cluster bootstrap uncertainty, limited nested parameter selection and feature ablations. Target-derived inputs increased coalfield-held-out LightGBM $${R}^{2}$$ from 0.383 to 0.920 for maximum horizontal stress, an absolute increase of 0.537. Full-feature Ridge achieved $${R}^{2}$$ of 0.440 and 0.425 for the two horizontal stresses; depth-only models reached 0.526 and 0.521. Model rankings depended on group weighting and validation partition. Conditional completion retained grouped $${R}^{2}$$ of 0.867–0.904. Group-block calibration increased record-level interval coverage to 97.9–99.3%, but substantially widened the intervals. In Lu'an, conditional completion achieved $${R}^{2}$$ near 0.93 under mine-group and representative-coordinate hold-outs, whereas blind transfer remained limited. Combining local and other-coalfield records improved hidden-point prediction in Lu'an and Xishan but not consistently in Jincheng. These findings support explicit information boundaries and group-specific evaluation for stress-database appraisal, while limiting claims of geographic transfer and engineering substitution.
Authors: Yi Li, Zhanshan Shi, Gang Li, Beifang Wang
Institution: Liaoning Technical University
Language: English
Subfield: Mechanics of Materials
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,890
OpenAlex ID: W7234195967
