Miners' Shift Reports as a Biased Sensor for Floor Heave

In deep mines and tunnels, sideways rock pressure often exceeds the weight of the rock above. Under those conditions the excavation floor can buckle upward, a failure called floor heave, more than the roof sags. In one near-fault tunnel simulation, floor heave reached 6.93 mm against 5.35 mm of vault (roof) settlement (Qingyun Huang et al. (2023)). The instruments that measure this are few and fixed in place. A roadway several kilometres long may have three monitoring stations, and campaigns typically last 30 to 40 days (Xiaoxu Gao et al. (2026); Zhongzong Cao et al. (2024)). Meanwhile, every shift, people walk the full length of the roadway and write down what they saw. Those shift reports form a dense record covering the whole roadway over long periods. Nobody uses them quantitatively, partly because self-reports are known to be biased in systematic ways.

This proposal treats shift reports as a sensor whose errors can be measured and modelled, rather than as either truth or noise. The working hypothesis is that the errors have a structure. Their size may depend on where in the opening the deformation is: roof, floor, or sidewall, since a sagging roof is a more vivid threat than a slowly rising floor. It may depend on how long after the observation the report is written. It may also depend on how dangerous the zone felt. If that structure can be estimated against instrumented sections and then corrected for, debiased reports can fill the gaps between gauges. The intended payoff is better estimates of floor heave between instruments, which is the deformation the hypothesis expects people to under-report.

Core Claim

Human deformation reports are a component-specific, delay-dependent measurement of rock movement. A small measurement-error model makes this concrete. The model is calibrated on sections where laser scans and gauges give ground truth. It is then combined with the sparse gauges along the roadway. On held-out sections, the combined estimate of floor heave should be more accurate than either gauges alone or reports taken at face value. The bias terms are scientifically interesting in their own right, beyond their engineering use. The proposal states the salience hypothesis (people over-report the roof and under-report the floor) as a hypothesis, not a known fact, and designs the study so it can fail.

Each piece of prior work plays a specific role here. The mechanics literature motivates the target. It shows that floor and vault deform unequally (Qingyun Huang et al. (2023)). It explains why: the roof can lock into a self-supporting arch of fractured blocks (a "voussoir beam") while the unconfined floor yields (Xufeng Wang et al. (2017)). It also shows that how unequal the deformation is depends on site and stress, through the ratio of horizontal to vertical stress at which the failed-rock zone switches to a "butterfly" shape (Xiaoyong Lian et al. (2023)).

The recall literature supplies the error mechanisms we parameterize, but it studies health and life events rather than physical deformation. First, a person's current state recolours what they remember (Wenqing Wu et al.; John R. Pfeiffer et al.). Second, after an incident people inflate how good things were beforehand (Ronan A Lyons). Third, memories of how often something happened decay with time (M. Zhang et al.).

The self-report literature supplies the method of validating against objective measures. Examples are biomarker checks of recalled drug use (Pengfei Guo et al. (2025)) and passive sensing replacing stress surveys (Tiantian Feng and Shrikanth Narayanan (2024)). It also identifies social desirability, meaning answers bent to look good, as a separate distortion (Heng Yue et al.). In mining, social desirability plausibly takes the form of under-reporting to avoid halting production.

The research file found no work connecting rock deformation to human reporting. The closest methodological relative is FLARE (Ruxiao Chen et al. (2025)). FLARE uses a language model guided by behavioural theory to predict wildfire evacuation decisions from post-disaster surveys. It never compares the survey answers against physical ground truth, so it can model what people decided but cannot measure how wrong their recollections are. Our design does the opposite: the physical measurement is the target, and the human report is the biased input.

Research Program

The study runs at instrumented roadways, chosen so that the physical asymmetry differs between sites. The Yanghuopan return-air roadway in a multi-seam "isolated island" layout passes through five stress zones: two relieved zones beneath old worked-out areas, two concentrated zones beneath leftover coal pillars, and one undisturbed zone (Xiaoxu Gao et al. (2026)). The Yushuling roadways have the opposite ordering, with floor deformation the smallest: left side > roof > right side > floor (Zhongzong Cao et al. (2024)). Where possible we add a heave-dominated tunnel section like the F1 fault case. This contrast is the core of the design. If report bias tracks which surface is visually threatening and stays the same while the physical ordering flips, that is evidence of a perceptual bias. If reports simply follow magnitude, the salience hypothesis is wrong.

Ground truth comes from existing gauges and bolt-force sensors at stations like K1–K3, plus weekly laser scans of the roadway surface along calibration segments. Gao et al. explicitly call for remote sensing because point gauges miss continuous floor movement, and the scans answer that call. Human reports use a short structured form for each 10 m segment. It asks for a 0–4 rating of roof, floor, and each sidewall, plus free text.

Two factors are randomized within workers. The first is timing: the form is filled in at the segment, or at the end of the shift. This manipulates retrospective delay directly, borrowing the micro-randomized design of Jessica R. Golbus et al. (2026). The second is attribution: the report is anonymous or signed, which isolates social desirability. Exposure to danger is not randomized. Instead it is coded from the mechanics, using whether the segment lies within the 30–38 m zone of raised pressure ahead of the mining face (Xiaoyong Lian et al. (2023)) or within a pillar's 62.1 m stress shadow.

The measurement model links report $r$ to true deformation $u$:

$$ r_{w,l,s,t} = a_l + b_l\,u_{l,s,t} + \gamma_l\,d + \delta_l\,z_s + \eta_w + \varepsilon . $$

Here $w$ is the worker, $l$ the surface (roof, floor, or a sidewall), $s$ the segment, and $t$ the shift. The terms are defined as follows. $a_l$ and $b_l$ are each surface's offset and sensitivity. $\gamma_l$ captures the effect of reporting delay $d$, which is either in-place or end-of-shift. $\delta_l$ captures the effect of the high-stress zone indicator $z_s$. $\eta_w$ is a per-worker random effect. $\varepsilon$ is residual noise.

The salience hypothesis predicts that floor sensitivity is lower than roof sensitivity, $b_{\text{floor}} < b_{\text{roof}}$, after controlling for actual magnitude. Inverting the model gives a debiased deformation estimate with its own uncertainty. That estimate enters a Gaussian-process (Kriging-style) interpolation along the roadway together with the gauges. Spatial Kriging is already used for gas-hazard prediction in coal seams (Qingsong Li et al. (2025)), and here it gains a second, noisier data source.

Baselines

The naive reference uses gauges only: the Gaussian-process interpolation along the roadway with no human input. This is what a mine has today. Every other method must beat it on held-out laser-scanned segments, using the same kernel family and the same tuning budget.

The second baseline takes reports at face value. Shift ratings are converted to millimetres with a single global linear fit and fused with the gauges. This tests whether reports carry any usable signal at all. It also serves as the main ablation: if it matches the full model, the surface-specific and delay terms add nothing.

The third baseline is a constant-bias correction: one offset and slope shared across all surfaces, with no delay or stress-zone terms. It isolates the key ingredient, which is surface-specific bias. Beating it is the result that justifies the proposal's central claim.

The strongest existing method is a geotechnical machine-learning regressor of the kind surveyed by Wei Shao et al. (2023). It predicts deformation from stratigraphy, stress zone, and distance to the face, with no human input. That review reports support vector machines (SVM) reaching a regression coefficient of $R^2 = 0.99$ for rock deformation prediction. That figure comes from curated datasets of a different task, so we do not treat it as the bar on our data. For human-report-informed roadway deformation estimation, no reported result was found on Lacuna. We train the SVM and a random forest on the same calibration segments with nested cross-validation and the same number of hyperparameter trials. We also give them the gauges as features, so the comparison isolates what the reports add.

The fifth baseline is a language-model extractor in the spirit of Ruxiao Chen et al. (2025). It reads the free text and outputs deformation estimates per surface, using structured step-by-step prompting with no measurement model. FLARE's own comparison, 89.5% for the theory-guided system versus 73.8% for an unguided language model on evacuation prediction, warns that raw extraction is weak. It does not tell us how well extraction works on deformation.

Beating the baselines means two things on held-out segments at both sites. Floor-heave error must be lower than every baseline, with confidence intervals that exclude zero difference. Recall for detecting segments whose heave exceeds a support-intervention threshold must also be higher.

Why not just…?

Why not just install more sensors or scan continuously? This is the real threat. If continuous laser scanning of the whole roadway becomes cheap, human reports become redundant for estimation. Floor sensors are exposed to traffic and to heave itself. Field campaigns in the file are short and sparse (Xiaoxu Gao et al. (2026)), and the scans in this study cover only calibration segments weekly. We measure the threat directly. We plot fused-estimate error against scan density, and report the density at which reports stop helping. If that density is already affordable, we say so, and the contribution shrinks to the bias science.

Why not just use the geotechnical model without people? Physics-informed regressors encode where heave should occur, but site-specific thresholds do not transfer between stratigraphies (Xiaoyong Lian et al. (2023)). Reports observe what is actually happening in each segment. The SVM baseline, given the same gauges, settles whether that observation adds information.

Why not just have workers report in place and skip recall modelling? In-place reporting removes delay but not salience or social desirability. Momentary reports in the residents' study still left modest predictive performance, with a macro-F1 of 56.20 to 63.41 (Tiantian Feng and Shrikanth Narayanan (2024)). The randomized timing arm answers this directly. If $\gamma_l$ is near zero and the surface-specific terms vanish under in-place reporting, the simpler protocol wins, and we would recommend it.

Decisive Checks

The first check is the cross-site sign test. The sensitivity gap between roof and floor, $b_{\text{roof}} - b_{\text{floor}}$, should have the same sign at Yanghuopan and Yushuling even though the physical ordering of deformation differs. A gap that flips with physical magnitude falsifies the salience hypothesis.

The second check is the held-out fusion test. Calibrate at one site, then apply the model to new segments and new workers at the other. Floor-heave error should beat the constant-bias ablation. Calibrating per site should be required only for $a_l$, not for $b_l$.

The third check is the state-dependence test. If $\delta_l$ is nonzero inside the pressure zone ahead of the face and grows under end-of-shift reporting, that would be the first evidence that dangerous ground conditions recolour retrospective hazard reports.

The fourth check is the anonymity contrast. Lower floor ratings in signed reports, at matched true deformation, would separate production-protecting under-reporting from perceptual bias. That distinction changes what a mine should do about the reports.

References

  1. Mechanical Response Characteristics Analysis of Tunnel Surrounding Rock during Near-Fault Construction
  2. Research on stability analysis and control technology of roadways in the underlying isolated island working face
  3. Surrounding Rock Control Technology of Thick Hard Roof and Hard Coal Seam Roadway under Tectonic Stress
  4. Characteristics of Strata Behavior during Thick Seam Mining by Fully-Mechanized Top Coal Caving in a Loess-Covered Gullied Region
  5. Law of strata pressure behavior of surrounding rock in nearby goaf roadway for extra-thick coal seams of Datong mine area
  6. Clinical Features of Patients with Dysthymia in a Large Cohort of Han Chinese Women with Recurrent Major Depression
  7. Associations between childhood family emotional health, fronto-limbic grey matter volume, and saliva 5mC in young adulthood
  8. Measuring the Population Burden of Injuries—Implications for Global and National Estimates: A Multi-centre Prospective UK Longitudinal Study
  9. Alternative risk factors of HCV infection in a rural community in China
  10. Acetaminophen Use in Pregnancy: A Comparison of Self-Reported Intake with Maternal and Newborn Biomarker Measures
  11. Understanding Stress, Burnout, and Behavioral Patterns in Medical Residents Using Large-scale Longitudinal Wearable Recordings
  12. Measurement Invariance of the Bergen Social Media Addiction Scale Across Genders
  13. From Perceptions to Decisions: Wildfire Evacuation Decision Prediction with Behavioral Theory-informed LLMs
  14. Impact of Push Notifications on Physical Activity and Sodium Intake Among Patients with Hypertension: Microrandomized Trial of a Just-in-Time Adaptive Intervention
  15. Spatial Risk Prediction of Coal Seam Gas Using Kriging Under Complex Geological Conditions
  16. The Application of Machine Learning Techniques in Geotechnical Engineering: A Review and Comparison