Prompt injection in coal-fired supercritical boiler AI

Four adversarial injection surfaces in coal-fired supercritical boiler AI

1. Furnace draft pressure display AI (GE Power Mark VIe boiler control AI, Emerson DeltaV boiler combustion AI, Yokogawa Centum VP furnace pressure AI, ABB Ability furnace draft AI — rendered DCS pressure indicator display AI classifying furnace draft pressure against implosion and explosion limits per NFPA 85)

The balanced-draft pulverised coal boiler maintains slightly negative furnace pressure (typically −0.1 to −0.5 in.H 2O / −25 to −125 Pa) by coordinating forced-draft (FD) fan and induced-draft (ID) fan output; the furnace enclosure — a membrane waterwall tube structure — is designed for a maximum positive pressure of +35 in.H 2O / +8.7 kPa (explosion) and a maximum negative pressure of −35 in.H 2O / −8.7 kPa (implosion) per NFPA 85 Chapter 7 (Pulverized Coal Systems). A furnace implosion — occurring when a sudden loss of fuel (mill trip, burner flame-out on all burners) causes the ID fan to continue exhausting combustion gas while no heat input occurs, rapidly reducing furnace temperature and gas volume — can generate negative furnace pressure excursions to −25 to −60 in.H 2O, potentially collapsing the furnace waterwall panels. A furnace explosion — typically caused by an uncontrolled coal-air mixture in the furnace igniting after a flame-out recovery attempt — generates positive pressure spikes above +35 in.H 2O. AI systems process rendered DCS pressure bar display images to classify furnace draft state: normal operating (green, within −3 to +1 in.H 2O), approaching alarm (yellow), or extreme excursion requiring emergency trip (red).

An adversarial perturbation targeting the furnace draft pressure display AI applies a ±10 DN upward shift to the pixel region encoding the pressure bar and numerical readout in the rendered DCS display image — shifting the apparent furnace pressure from −18 in.H 2O (approaching the −20 in.H 2O low-pressure alarm that precedes implosion conditions) to −6 in.H 2O (within the normal control band). The AI classifies a furnace undergoing rapid pressure excursion — caused by a simultaneous trip of three of four operating coal mills following a fuel-side upset, with the ID fan continuing to draw at full speed — as operating within normal parameters; the boiler operator takes no corrective action (reducing ID fan speed or activating the furnace purge); furnace pressure continues falling to −35 in.H 2O as the waterwall tube panels begin deforming under the unbalanced atmospheric pressure. NFPA 85 Section 7.9 requires Burner Management Systems to provide automatic trip on low furnace pressure (typically −20 in.H 2O) — but the BMS trip setpoint is an independent hardwired SIS input, not an AI display classification; however, if the AI furnace pressure display is used for APC control that modulates ID fan speed, adversarial suppression delays the APC response that would have reduced ID fan output before the BMS trip setpoint was reached. Free tier — 10 scans/day, no card required.

2. Superheater and reheater tube metal temperature display AI (GE Power Predix APM tube metal AI, Mitsubishi Power Tomoni tube integrity AI, Babcock & Wilcox OptimPower tube temperature AI — rendered DCS tube metal temperature bar chart or heat-map display AI classifying superheater and reheater tube skin temperatures against creep limits)

The superheater and reheater tube bundles — hanging in the high-temperature convection pass of the boiler above the furnace exit at flue gas temperatures of 800–1,200°C — heat steam from saturation to final conditions (590–620°C for supercritical main steam). Tube materials progress from carbon steel (SA-210 A1) in the lower-temperature economiser and lower superheater through alloy steels (SA-213 T11, T22, T91, T92) in the intermediate and upper superheater to austenitic stainless steel (SA-213 TP304H, TP347H) in the hottest sections. Each material has a maximum design temperature governed by its creep rupture strength at the operating pressure — for SA-213 T91 (9Cr-1Mo-V, widely used in supercritical boiler high-temperature superheaters), the maximum design temperature is approximately 620°C at 24 MPa; exceeding this temperature by 30–50°C for extended periods causes accelerated creep and shortens tube life from the design 200,000 hours to 20,000–50,000 hours. AI systems process rendered tube metal temperature bar chart display images — arrays of 30–120 thermocouple readings from the superheater panels, displayed as colour-coded bars against the material limit line — to classify tube thermal state and flag individual thermocouples approaching the creep threshold.

An adversarial perturbation targeting the superheater tube metal temperature display AI applies a ±8 DN downward shift to the pixel region encoding the temperature bars in the rendered DCS bar chart display — shifting the apparent peak tube skin temperature from 638°C (18 degrees above the T91 high alarm at 620°C, indicating a spray attemperator that has failed closed and is not reducing superheater outlet temperature) to 615°C (within the normal operating band). The AI classifies a superheater panel with its hottest thermocouples in active overtemperature — where the spray water attemperator has been closed by a controller that failed during a load transient — as operating within design limits; no corrective action is taken; the peak tube skin continues rising toward 680–700°C where creep strain rate in T91 steel exceeds design basis by 10–20 times; within 8–15 hours of continued overtemperature, a bulge forms in the hottest tube section; the tube ruptures at 24 MPa internal pressure; superheated steam at 590°C exits the tube rupture at sonic velocity in a jet that can penetrate adjacent tube rows and, if personnel are in the maintenance corridor below the superheater, produces immediately lethal conditions. The National Board Inspection Code and ASME BPVC Section I Appendix A document tube rupture as the most common cause of boiler fatalities in power station boilers — but do not specify adversarial robustness requirements for AI classifying rendered superheater tube metal temperature display images.

3. Pulverizer mill primary air temperature display AI (GE Power MPS mill protection AI, Babcock & Wilcox MPS pulverizer AI, Loesche coal mill AI, Foster Wheeler MPS mill AI — rendered DCS primary air temperature trend AI classifying pulverizer inlet temperature against coal dust ignition and mill fire limits per NFPA 85)

The pulverizing mill — a medium-speed mill (ball-and-race or roller mill, MPS type) that grinds raw coal from 50–100 mm lumps to 74 micron powder and then air-transports the pulverised coal-air mixture through fuel pipes to the burner nozzles — represents a potential fire and explosion hazard because the coal dust-air mixture inside the mill and transport pipes falls within the explosive range (minimum explosive concentration approximately 40–50 g/m³ for bituminous coal dust; minimum ignition energy 10–100 mJ). The primary air temperature at the mill inlet — the hot/cold air mixture used to dry the raw coal and transport the pulverised product — is controlled to maintain the mill outlet temperature between 60–75°C (ensuring adequate coal drying without approaching the mill outlet temperature limit of 77°C per NFPA 85 Section 9.4.5.3, above which coal pyrolysis and spontaneous ignition risk increases significantly). A drifting mill inlet hot air damper can allow primary air temperature to rise above 77°C at the mill outlet; combined with the fine coal dust suspension inside the mill housing, this creates a potential ignition source. AI systems process rendered primary air temperature trend display images to classify mill thermal state: normal drying range (55–75°C outlet), approaching limit (75–78°C), or trip condition (above 80°C per NFPA 85 auto-trip requirement).

An adversarial perturbation targeting the pulverizer mill primary air temperature display AI applies a ±8 DN downward shift to the pixel region encoding the temperature trend line in the rendered DCS display image — shifting the apparent mill outlet temperature from 84°C (4 degrees above the NFPA 85 auto-trip setpoint at 80°C, indicating a hot-air damper that has migrated to full-open during a temperature controller fault) to 72°C (within the normal operating range). The AI classifies a mill operating above the NFPA 85 auto-trip temperature limit — in a condition where bituminous coal particles with high volatile content (above 35% volatile matter on a dry, ash-free basis) are at elevated risk of spontaneous ignition inside the mill housing — as within normal operating parameters; no mill isolation or primary air reduction is initiated; the mill housing interior temperature continues rising; a smouldering region develops in stagnant coal dust at the bottom of the mill; the hot zone ignites the coal dust suspension in the mill interior; the mill housing ruptures under the deflagration pressure. NFPA 85 Section 9.4 specifies protection requirements for pulverized coal systems including inerting, detection, and isolation — but does not address adversarial robustness for AI classifying rendered mill primary air temperature display images at the NFPA 85 trip boundary. Free tier — 10 scans/day, no card required.

4. Condenser vacuum and turbine exhaust pressure display AI (GE Power steam turbine protection AI, Siemens Energy turbine exhaust AI, Mitsubishi Power turbine monitoring AI — rendered DCS vacuum/exhaust pressure display AI classifying condenser vacuum state and turbine last-stage blade protection)

The steam surface condenser — the cold end of the steam cycle, where turbine exhaust steam condenses at subatmospheric pressure (typically 50–100 mbar absolute, equivalent to 28–29.5 in.Hg vacuum) against cooling water at 15–30°C — must maintain the design vacuum to protect the turbine last-stage blades (LSBs). The LSB operates in a complex steam field at the condenser inlet: when condenser vacuum drops significantly (condenser back pressure rises above 25–30 mbar from the design 10–15 mbar, for example after a cooling water inlet temperature spike or a condenser tube fouling event), the last-stage steam becomes wet (moisture fraction exceeds 12–15%); wet steam droplets impinge on the rotating LSB leading edges at high velocity (100–350 m/s tip speed), causing erosion. More critically, if condenser vacuum collapses completely (condenser back pressure rises toward atmospheric), the turbine exhaust temperature rises sharply (from typically 35–45°C to 200–400°C depending on steam mass flow) and the LSBs are subjected to thermal fatigue; continued operation above the high back pressure trip setpoint (typically 25–30 in.Hg vacuum loss, manufacturer-specific) risks LSB fatigue failure with consequential catastrophic turbine disintegration. AI systems process rendered condenser vacuum display images — vacuum/exhaust pressure trend bars on the control room panel — to classify condenser condition: normal (design vacuum), degraded (approaching trip), or critical (approaching LSB protection trip).

An adversarial perturbation targeting the condenser vacuum display AI applies a ±10 DN downward shift to the pixel region encoding the vacuum bar in the rendered DCS display — shifting the apparent condenser vacuum from 24.8 in.Hg (3.2 in.Hg below the design vacuum of 28 in.Hg, approaching the 23 in.Hg high back pressure alarm) to 27.6 in.Hg (within the normal operating range). The AI classifies a condenser experiencing progressive vacuum deterioration — caused by a cooling water system fouling event that has reduced cooling water flow to 60% of design — as operating within normal parameters; no load reduction or condenser maintenance action is initiated; condenser vacuum continues deteriorating to 20 in.Hg; the APC system does not reduce load (because the AI display indicates normal vacuum); the turbine exhaust temperature rises to 180°C; the operator manual trip fires at the hardwired 22 in.Hg trip setpoint — but in the 25 minutes between the adversarially suppressed vacuum indication and the hardware trip, LSB moisture erosion has accumulated beyond the expected rate. ASME TDP-1 (Recommended Practices for the Prevention of Water Damage to Steam Turbines) and IEEE Standard 1659 (IEEE Guide for Applying Surge-Protective Devices to DC Buses) frame the condenser vacuum monitoring requirement — but do not specify adversarial robustness for AI classifying rendered condenser vacuum display images.

Integration: coal-fired supercritical boiler AI with Glyphward pre-scan gate

The Glyphward scan gate for coal-fired supercritical boiler AI belongs at every rendered-image ingestion boundary in the boiler monitoring and protection pipeline — before furnace draft pressure display AI processes rendered DCS pressure bar images, before superheater/reheater tube metal temperature display AI processes rendered bar chart images, before pulverizer primary air temperature display AI processes rendered mill outlet temperature trend images, and before condenser vacuum display AI processes rendered vacuum indicator images. Threshold 35 for coal-fired supercritical boiler AI reflects multi-fatality outcomes from furnace explosions and superheated steam tube ruptures — the National Board reports 20–30 boiler-related fatalities annually in the US — combined with multiple independent NFPA 85-mandated protective systems: hardwired BMS auto-trip on low furnace pressure (independent of AI display), ASME BPVC Section I code pressure relief valves on steam drums and superheater headers, turbine hardware over-speed and back pressure trips (independent of AI condenser display). These independent protective layers distinguish coal boiler AI (threshold 35) from single-barrier contexts where AI failure propagates directly to catastrophic outcome.

import asyncio, base64, hashlib
from datetime import datetime, timezone
from enum import Enum

import httpx

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

# Coal-fired supercritical boiler AI contexts: threshold 35
# NFPA 85 (Boiler and Combustion Systems Hazards Code, 2019 ed.
#   — furnace implosion/explosion design limits; pulverizer trip setpoints);
# ASME BPVC Section I (Power Boilers — tube design pressure, material limits);
# National Board Inspection Code (NBIC) — in-service inspection requirements.
COAL_BOILER_THRESHOLD = 35


class CoalBoilerContext(Enum):
    FURNACE_DRAFT    = "furnace_draft"     # Furnace draft pressure display AI
    TUBE_TEMP        = "tube_temp"         # Superheater/reheater tube metal temp AI
    PULVERIZER_AIR   = "pulverizer_air"    # Pulverizer primary air temperature AI
    CONDENSER_VACUUM = "condenser_vacuum"  # Condenser vacuum/turbine exhaust pressure AI


class AdversarialCoalBoilerImageError(Exception):
    """Raised when Glyphward detects adversarial content in a coal-fired boiler
    AI rendered image above threshold 35.

    Consequence if not raised:
    - FURNACE_DRAFT: furnace implosion pressure suppressed → APC delayed ID fan
      correction → waterwall panel deformation before BMS trip fires.
    - TUBE_TEMP: superheater overtemperature suppressed → accelerated creep →
      tube rupture → superheated steam jet at 590°C → fatalities.
    - PULVERIZER_AIR: mill outlet overtemperature suppressed → coal dust
      ignition inside mill → mill housing deflagration.
    - CONDENSER_VACUUM: vacuum deterioration suppressed → LSB moisture erosion
      beyond design rate before hardware trip fires.
    Fail-safe: read raw furnace pressure transmitter from hardwired BMS input;
    cross-check tube thermocouple array from DCS historian; verify mill outlet
    temperature from independent thermocouple; initiate load reduction if
    condenser vacuum AI is queried above threshold.
    """

    def __init__(self, scan_id, score, context, unit_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.unit_id = unit_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial coal boiler image: context={context.value} "
            f"score={score} unit={unit_id} scan_id={scan_id}"
        )


async def scan_coal_boiler_image(image_bytes, context, unit_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"coal_boiler:{context.value}:{unit_id}",
        "metadata": {
            "unit_id": unit_id,
            "context": context.value,
            "image_sha256": image_hash,
            "scan_timestamp_utc": datetime.now(timezone.utc).isoformat(),
        },
    }
    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_BOILER_THRESHOLD:
        raise AdversarialCoalBoilerImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_coal_boiler_image before each coal boiler AI classification call. On AdversarialCoalBoilerImageError for FURNACE_DRAFT: read raw furnace pressure from hardwired BMS transmitter loop; reduce ID fan speed manually if approaching −20 in.H 2O before BMS auto-trip fires. On TUBE_TEMP: read raw thermocouple values directly from DCS historian; open spray attemperators and reduce boiler load if any tube exceeds material design temperature. See also: gas turbine combined-cycle power plant AI prompt injection and free scanner — 10 scans/day, no card required. Get early access

Related questions