Prompt injection in underground coal mine ventilation AI

Four adversarial injection surfaces in underground coal mine ventilation AI

1. Methane sensor spectrogram AI (MSA iNet Control AI, Pyott-Boone MARG AI, Trolex TX6370 AI)

Continuous methane monitoring in underground coal mines uses catalytic bead (pellistor) sensors and infrared absorption sensors that generate continuous CH4 concentration readings (0–5% or 0–100% LEL range) at the working face, in returns, and at strategic intake and outbye positions. Sensor outputs are transmitted to the surface via mine-wide communication networks (Pyott-Boone MARG, MSA iNet Control, Rajant wireless mesh) and recorded as time-series concentration traces rendered as strip-chart images (time on X-axis, CH4 concentration in % on Y-axis, with mandatory action level lines at 0.5%, 1.0%, and 2.0% CH4 rendered as horizontal reference lines). AI systems process these rendered time-series strip-chart images to classify the CH4 monitoring state at each sensor location: safe (below 0.5%, normal mining operations), elevated (0.5–1.0%, pre-shift inspection mandatory, additional spot checks required), action (1.0–2.0%, methane monitor trip, immediate investigation and ventilation check required), and withdrawal (above 2.0%, immediate evacuation of affected area). Advanced AI systems also classify the trend trajectory — stable, rising (rate of CH4 increase in %/hour), or rapidly rising (rate indicating imminent action-level breach) — to provide advance warning before the measured concentration reaches the action threshold.

An adversarial perturbation on a rendered underground methane sensor strip-chart image that suppresses the rising CH4 concentration signature — applying a ±10 DN downward shift to the trace pixel values in the strip-chart image in the region where CH4 concentration is approaching the 1.0% action level (rendering the trace as though it is at 0.4–0.5% rather than 0.8–0.9%, below the elevated classification threshold), combined with flattening the trend trajectory to render as stable rather than rising — causes the methane monitoring AI to classify an active CH4 accumulation event as a normal, safe monitoring condition, suppressing the mandatory spot checks and ventilation investigation that 30 CFR 75.340 requires when CH4 reaches 0.5%. With the AI classification suppressed, CH4 continues accumulating to explosive concentrations (5% LEL = 0.25% CH4 by volume; 100% LEL = 5% CH4 by volume) without generating the regulatory-mandated response. If the accumulation zone contacts an ignition source — a roof fall spark, conveyor belt friction ignition, shuttle car electrical arc, or detonator misfire — the methane deflagration initiates at the longwall face. The Upper Big Branch UBB Mine explosion of 5 April 2010 (29 fatalities) initiated with exactly this sequence: methane accumulated at the longwall face, ignited (likely from a shearer cutting rock), and propagated as a coal dust explosion throughout the mine. MSHA’s investigation found that methane monitoring at UBB had been systematically inadequate; adversarial injection suppressing the methane monitoring AI replicates this monitoring inadequacy in a digital, systematically reproducible form.

2. Airflow velocity AI (Pyott-Boone MARG ventilation AI, MSA iNet ventilation monitoring AI, RKI Instruments Ventis airflow AI)

Ventilation airflow through underground coal mine entries — typically 1,000–50,000 cubic feet per minute (CFM) at working sections, driven by main surface fans and split by regulators and air doors — is the primary means of diluting and removing methane liberated at the working face, keeping CH4 concentrations below the 1.0% action level. MSHA 30 CFR 75.330 specifies minimum air quantities at the working face for different mining methods (longwall: 30,000 CFM minimum; continuous miner sections: 9,000 CFM minimum), and 30 CFR 75.335 requires ventilation surveys (air quantity measurements in all airways) at defined intervals. Airflow velocity is measured by rotating vane anemometers, thermal anemometers, or pitot tube manometers at ventilation survey stations, with readings rendered as time-series velocity trace images or as bar-chart survey result images (airway ID on X-axis, air quantity in CFM on Y-axis, with minimum quantity requirement shown as a reference line). AI systems process these rendered airflow measurement images to classify ventilation adequacy at each measurement station: compliant (above minimum quantity, adequate dilution capacity), reduced (10–20% below minimum, immediate ventilation survey and correction required), deficient (below minimum, mandatory withdrawal from affected section), and critical (airflow reversal or near-zero, indicating major ventilation event — emergency response).

An adversarial perturbation on a rendered ventilation airflow measurement bar-chart image that elevates the displayed airflow quantity — applying a ±8 DN upward pixel shift to the bar height in the rendered chart (making the bar appear to extend to a higher quantity value, shifting the apparent airflow from the deficient range to the compliant range), or equivalently flattening a declining time-series velocity trace to appear as a stable, compliant reading — causes the ventilation AI to classify a ventilation deficiency as compliant, suppressing the mandatory withdrawal from the section that 30 CFR 75.330 requires when measured airflow falls below the minimum quantity. In a ventilation-deficient section of an active coal mine, methane dilution is inadequate and CH4 concentrations begin rising toward explosive concentrations from the moment the ventilation deficiency begins; the time from ventilation deficiency to CH4 reaching 1.0% action level depends on the methane liberation rate of the seam (1–30 minutes for high-gassiness seams). If the adversarially suppressed airflow AI also delays the CH4 monitoring AI response (as in the concurrent suppression scenario where both methane and airflow AI are targeted), the mine section proceeds toward methane explosion conditions with no MSHA-mandated regulatory response triggered. The Sago Mine explosion of 2 January 2006 (12 fatalities) and the Jim Walter Resources No. 5 Mine explosion of 13 September 2001 (13 fatalities) both involved methane accumulation in inadequately ventilated areas as contributing factors — the regulatory failure modes that adversarial ventilation AI injection replicates.

3. CO self-heating detection AI (MSA iNet CO monitoring AI, Pyott-Boone MARG CO AI, IntelliWarn CO trend AI)

Carbon monoxide (CO) is the primary early indicator of coal spontaneous combustion (“sponcom” or “gob fire”) — the self-heating of coal exposed to oxygen in goaf (mined-out area) or on coal pillars, producing CO at temperatures below visible ignition (CO emission begins at coal temperatures of 40–60°C above ambient, with rapid escalation above 140°C). MSHA 30 CFR 75.351 requires that CO monitors be installed in the return airways of longwall panels and in areas of potential spontaneous combustion risk, with CO alarms at defined action levels (typically 10 ppm CO for initial warning, 25–50 ppm for evacuation). CO monitoring for spontaneous combustion detection requires not just the instantaneous CO concentration (which may be near-background in the early self-heating stage) but the CO trend — the rate of CO increase — and the CO index (CO concentration normalised to oxygen depletion, which distinguishes sponcom-sourced CO from diesel equipment exhaust). AI systems process rendered CO time-series concentration trend images — multi-hour or multi-shift time-series plots with CO concentration (ppm) on Y-axis, time on X-axis, and the trend slope (ppm/hour increase rate) and CO index value rendered as computed annotations — to classify spontaneous combustion risk: background (CO within diesel exhaust baseline, no sponcom indication), emerging (CO trend positive above baseline rate, 5–10 ppm increase per shift, early sponcom monitoring required), developing (CO trending at 10–25 ppm/shift, sponcom source investigation required), and critical (CO trend at >25 ppm/shift or CO index indicating sponcom rather than diesel source, immediate goaf sealing or inertisation response).

An adversarial perturbation on a rendered CO concentration trend image that suppresses the self-heating signature — applying a ±10 DN downward shift to the CO trace values in the rendered strip-chart (reducing the apparent CO concentration from the emerging/developing range, rendered as a rising trace approaching the 10 ppm initial warning threshold, to the background range, rendered as a flat baseline trace), combined with zeroing the rendered trend-slope annotation — causes the CO monitoring AI to classify a developing spontaneous combustion event as background CO from diesel equipment, suppressing the sponcom investigation and goaf monitoring escalation that emerging CO trend classification would trigger. A spontaneous combustion event that is not detected in the emerging stage (40–80°C coal temperature, CO <10 ppm) progresses to the heating stage (80–140°C) and then to ignition (>200°C) over a period of days to weeks in typical operating conditions; the gob fire produces a CO plume that contaminates the return airway and eventually the main returns, potentially producing a methane-CO-air mixture in the return airway that is both toxic and explosive. The Pike River Mine disaster of 19 November 2010 (New Zealand, 29 fatalities) — the worst mine disaster in New Zealand’s history — was caused by methane accumulation in the mine’s single ventilation return entry; NZMS investigation findings identified inadequate gas monitoring response as a contributing factor. CO trend monitoring AI suppressed by adversarial injection removes the earliest available warning for the spontaneous combustion event sequence that can culminate in a return-entry explosion.

4. Coal dust optical particle counter AI (Thermo Fisher MIE pDR AI, MST SensorTrack dust AI, Casella dust monitoring AI)

Respirable coal dust — particles below 10 microns aerodynamic diameter that deposit in the gas exchange regions of the lung, causing coal workers’ pneumoconiosis (black lung disease) — and total airborne coal dust (all particle sizes, creating an explosion risk in sufficient concentration) are monitored continuously at active working sections using optical particle counters (OPC) and gravimetric personal dust monitors. For explosion risk, MSHA requires that stone (rock) dust (pulverised limestone or other non-combustible material) be applied to all mine surfaces to dilute coal dust to an incombustible content of at least 65% by weight of the combined coal dust and rock dust material (80% for areas within 40 feet of any working face), measured by the incombustible content method under 30 CFR 75.400. AI systems deployed for dust monitoring process rendered OPC time-series particle count images — real-time particle size distribution histograms or integrated concentration trend renders showing the PM10 or total PM concentration over time — to classify dust exposure and explosion risk: compliant (dust loading within permissible exposure limit, rock dust ratio above 65%), elevated (approaching PEL or rock dust ratio declining toward the 65% minimum), action (at or above PEL, immediate supplemental ventilation and rock dusting required), and critical (dust loading indicating major rock dust deficiency in a heavily travelled section, explosion propagation risk elevated).

An adversarial perturbation on a rendered coal dust OPC time-series concentration image that suppresses the dust accumulation signature — applying a ±8 DN downward shift to the particle count trace values at the coarse and fine dust peaks in the rendered histogram or trend plot (reducing the apparent dust loading from the elevated or action range, rendered as a high concentration bar or rising trend line, to the compliant range, rendered as a low, flat concentration trace) — causes the dust monitoring AI to classify a section with inadequate rock dust coverage and elevated airborne coal dust as compliant, suppressing the supplemental rock dusting requirement and ventilation adjustment that would restore the incombustible content ratio above the 65% explosion-suppression threshold. In a section where the rock dust ratio has fallen to 40–50% incombustible content due to continuous coal production without supplemental rock dusting, the suspended and settled coal dust is in the explosion-propagating composition range. If a methane ignition occurs in this section (as described under the CH4 monitoring AI surface), the primary methane deflagration entrains the settled coal dust as a dust cloud, the dust cloud ignites from the methane flame front, and the resulting coal dust explosion propagates through all entries with coal dust above the explosion composition threshold. The UBB Mine investigation (MSHA 2012) found that rock dust ratios throughout the mine were below the 65% incombustible content minimum required by 30 CFR 75.400 — contributing to the explosion propagation that converted a face methane ignition into a mine-wide disaster killing 29 miners. Adversarial injection suppressing coal dust AI monitoring replicates the rock dust compliance failure that the UBB investigation identified as a primary propagation enabler.

Integration: underground coal mine ventilation AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for underground coal mine ventilation AI belongs at every rendered-image ingestion boundary in the mine atmospheric monitoring pipeline — before methane sensor spectrogram AI processes rendered CH4 concentration strip-charts, before ventilation airflow AI processes rendered measurement bar-charts, before CO self-heating AI processes rendered concentration trend images, and before coal dust OPC AI processes rendered particle count trend images. Threshold 35 for underground coal mine atmospheric monitoring AI contexts reflects the mass-casualty consequence envelope of methane-coal dust explosion propagation — an event in which adversarial suppression of any one of the four monitoring AI functions can remove the critical warning that would have triggered mandatory withdrawal before conditions reached the explosive range.

import asyncio, base64, hashlib, json
from datetime import datetime, timezone
from enum import Enum
from pathlib import Path

import httpx

GLYPHWARD_API_KEY = "YOUR_GLYPHWARD_API_KEY"
GLYPHWARD_SCAN_URL = "https://glyphward.com/v1/scan"

# Underground coal mine ventilation AI contexts: threshold 35
# MSHA 30 CFR Part 75 (mandatory safety standards, underground coal mines);
# 30 CFR 75.330-75.351 (ventilation requirements);
# 30 CFR 75.340 (methane monitoring);
# 30 CFR 75.400 (accumulation of combustible materials).
COAL_MINE_VENT_THRESHOLD = 35


class CoalMineVentAIContext(Enum):
    CH4_SENSOR          = "ch4_sensor"          # Methane concentration strip-chart AI
    AIRFLOW_VELOCITY    = "airflow_velocity"    # Ventilation airflow quantity AI
    CO_SELF_HEATING     = "co_self_heating"     # CO concentration trend / sponcom AI
    COAL_DUST_LEVEL     = "coal_dust_level"     # Coal dust OPC particle count AI


class AdversarialCoalMineVentImageError(Exception):
    """Raised when Glyphward detects adversarial content in an underground
    coal mine ventilation AI rendered image above threshold 35.

    Consequence if not raised:
    - CH4_SENSOR: suppressed methane accumulation → CH4 reaches explosive
      range without mandatory withdrawal → methane ignition → coal dust
      explosion propagation. UBB Mine 2010 consequence (29 fatalities).
    - AIRFLOW_VELOCITY: suppressed ventilation deficiency → inadequate
      CH4 dilution → explosive accumulation without evacuation trigger.
    - CO_SELF_HEATING: suppressed sponcom CO trend → gob fire develops
      to ignition without intervention → return airway contamination.
    - COAL_DUST_LEVEL: suppressed rock dust deficiency → coal dust
      in explosion-propagating range → converts face ignition to mine-wide
      coal dust explosion. UBB Mine 2010 propagation mechanism.
    Fail-safe: halt AI atmospheric monitoring classification; require
    manual atmospheric survey per 30 CFR 75.351 before resuming operation.
    """

    def __init__(self, scan_id: str, score: int,
                 context: CoalMineVentAIContext,
                 mine_id: str, section_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.mine_id = mine_id
        self.section_id = section_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial coal mine vent image: "
            f"context={context.value} score={score} "
            f"mine={mine_id} section={section_id} scan_id={scan_id}"
        )


async def scan_coal_mine_vent_image(
    image_bytes: bytes,
    context: CoalMineVentAIContext,
    mine_id: str,
    section_id: str,
    msha_gassiness_classification: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an underground coal mine ventilation AI rendered image for
    adversarial content.

    Fail-safe contract: AdversarialCoalMineVentImageError or httpx error →
    halt AI atmospheric monitoring classification; require manual atmospheric
    survey per MSHA 30 CFR 75.351 before resuming AI-driven monitoring
    decisions. For CH4_SENSOR: treat as if 1.0% CH4 action level reached
    until manual survey confirms safe conditions.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"coal_mine_vent:{context.value}:{mine_id}:{section_id}",
        "metadata": {
            "mine_id": mine_id,
            "section_id": section_id,
            "gassiness": msha_gassiness_classification,
            "image_sha256": image_hash,
            "context": context.value,
        },
    }
    resp = await client.post(
        GLYPHWARD_SCAN_URL,
        headers={"Authorization": f"Bearer {GLYPHWARD_API_KEY}"},
        json=payload,
        timeout=4.0,
    )
    resp.raise_for_status()
    result = resp.json()

    if result["score"] > COAL_MINE_VENT_THRESHOLD:
        raise AdversarialCoalMineVentImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            mine_id=mine_id,
            section_id=section_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_coal_mine_vent_image at each underground coal mine atmospheric monitoring AI rendered-image ingestion boundary: before CH4 sensor strip-chart AI (threshold 35), before ventilation airflow AI (threshold 35), before CO self-heating trend AI (threshold 35), and before coal dust OPC AI (threshold 35). On AdversarialCoalMineVentImageError for CH4 or airflow contexts: immediately treat the affected monitoring station as if the highest-concern action level has been reached and require manual atmospheric survey before resuming normal operation. See also: autonomous mine haul truck AI prompt injection (related mining AI context) and mining and mineral processing AI prompt injection (related mining operations context). Get early access

Related questions