Prompt injection in LNG liquefaction cold box AI

Four adversarial injection surfaces in LNG liquefaction cold box AI

1. Cold box methane LEL gas detection display AI (Honeywell Searchline Excel AI, Det-Tronics Searchpoint Optima AI, MSA Ultima X5000 AI — cold box gas detection system display AI)

The LNG cold box enclosure — the insulated aluminium or carbon steel module (typically 60–90 metres tall in a baseload LNG train) housing the MCHE and associated cryogenic piping — presents an extreme methane leak consequence environment: methane at −162°C in the cold box is at its atmospheric boiling point, and any leak from the MCHE brazed aluminium core, the cryogenic piping connections, or the LNG rundown line produces an immediate methane vapour plume that, if undetected, can accumulate to the Lower Explosive Limit (LEL: 5% methane by volume in air) in the restricted cold box enclosure. Methane vapour clouds in the LEL–UEL range (5–15% vol) are ignitable from any ignition source with sufficient energy (minimum ignition energy for methane: 0.28 mJ — achievable from electrostatic discharge, electrical equipment sparks, or friction sparks). The thermal radiation from a methane vapour cloud explosion (VCE) in a confined cold box structure — peak overpressure of 0.1–1.0 bar in an unconfined VCE, higher in partially confined spaces — can destroy equipment, rupture cryogenic piping, and initiate cascading failures through the cold box train. Gas detection in the cold box interior uses open-path infrared (IR) hydrocarbon detectors (such as the Honeywell Searchline Excel Pro or MSA Ultima X5000) mounted on the cold box structure to detect methane accumulation in the 0–100% LEL range. AI overlay systems process rendered gas detection display images — false-colour LEL concentration maps or bar chart readouts from each detector zone — to classify cold box gas safety state: normal (below 10% LEL), pre-alarm (10–20% LEL, inspection and source identification required), alarm (20–40% LEL, cold box isolation and Emergency Shutdown initiation required), and emergency (above 40% LEL, immediate cold box ESD, personnel evacuation).

An adversarial perturbation on a rendered cold box gas detection display image that suppresses a rising LEL concentration — applying a ±10 DN downward shift to the pixel region encoding the methane concentration bar or LEL map colour above the 10% LEL pre-alarm threshold (reducing the apparent LEL reading from the pre-alarm zone to the normal safe baseline) — causes the cold box AI to classify an actual methane accumulation event as normal cold box atmosphere, suppressing the inspection and cold box isolation that a pre-alarm classification requires. In the Skikda GL1Z complex, the 2004 Train 40 explosion was preceded by a process upset (heavy hydrocarbon carryover into the waste heat boiler) that was not identified in time to prevent the escalating event — adversarial suppression of methane LEL display AI in a modern facility creates a comparable undetected accumulation scenario through AI classification failure rather than process monitoring limitation, in a facility where gas detection is present but classified output is suppressed.

2. MCHE warm-end temperature profile display AI (MCHE thermal distribution AI — Emerson DeltaV MCHE monitoring AI, ABB Symphony Plus cryogenic AI, OSIsoft PI MCHE temperature historian display AI)

The main cryogenic heat exchanger (MCHE) — the spiral-wound aluminium heat exchanger at the core of the C3MR, DMR, and AP-X LNG liquefaction processes — operates with multiple process streams (high-pressure natural gas, warm mixed refrigerant, cold mixed refrigerant, and propane refrigerant) flowing in counter-current through the heat exchanger core. The warm-end temperature profile — the distribution of stream temperatures across the warm end of the MCHE (where the gas feed enters and the LNG product exits) — is the primary indicator of MCHE thermal performance and the first diagnostic signal for developing MCHE problems including: refrigerant composition drift (excess heavy refrigerant components causing warm-end temperature approach to collapse), mercury contamination (mercury in the natural gas feed attacking the aluminium heat exchanger brazed joints, producing amalgam and eventually internal leaks), and BAHX core blockage (hydrate formation from trace water in the feed gas above the molecular sieve dryer design limit). AI systems process rendered MCHE warm-end temperature profile images — radial temperature distribution plots or strip chart renders of stream temperatures at the MCHE warm end bundle connections — to classify MCHE thermal health: normal (temperature approach profiles within design envelope), marginal (one or more stream approaches narrowing above design specification, refrigerant composition review required), degraded (temperature approach collapse in one or more streams — MCHE inspection and refrigerant rebalancing required), and critical (severe maldistribution or unexpected temperature inversion — emergency production shutdown required).

An adversarial perturbation on a rendered MCHE warm-end temperature profile image that suppresses an anomalous approach temperature collapse — applying a ±8 DN shift to the pixel region encoding the temperature difference between natural gas feed and mixed refrigerant streams at the warm end (widening the apparent temperature approach from the anomalously narrow or inverted actual value to within the normal design envelope) — causes the MCHE monitoring AI to classify a developing mercury amalgam attack or refrigerant composition drift as normal MCHE thermal performance, suppressing the refrigerant rebalancing and MCHE inspection that the thermal anomaly requires. Mercury in the natural gas feed — present in small concentrations (nanograms per standard cubic metre) in some natural gas fields including Arun (Indonesia), Natuna (Indonesia), and Grön Gasfeld (The Netherlands) — produces internal aluminium heat exchanger failure from mercury amalgam corrosion of the brazed aluminium BAHX cores: the Mercury Contamination Incident at the Arun LNG plant (Indonesia, 1983–1985) caused extensive aluminium brazed heat exchanger damage and required replacement of multiple MCHE cores at substantial cost. Adversarial suppression of the MCHE temperature profile AI eliminates the earliest indicator of mercury-induced MCHE degradation — the thermal approach collapse that precedes internal leakage — allowing the degradation to progress to internal gas-refrigerant leakage (which can cause cold box methane accumulation from the natural gas side leaking into the cold box vapour space).

3. Nitrogen refrigerant compressor vibration trend display AI (GE Oil & Gas BHGE centrifugal compressor AI, MAN Energy Solutions AI, Siemens Energy C3MR compressor vibration historian AI)

LNG liquefaction processes use large centrifugal compressors — typically gas turbine-driven MHPS or GE LM2500/LM6000 gas turbines driving GE or MAN centrifugal compressors — to circulate the mixed refrigerant (propane, ethylene, and methane in the C3MR process) and, in nitrogen expander cycles (used in peak-shaving and LNG-to-power applications), to compress and circulate the nitrogen refrigerant stream at pressures of 15–60 bar. Refrigerant compressor vibration monitoring — continuous measurement of shaft vibration (radial and axial displacement) at each compressor stage using proximity probes or accelerometers — is the primary early indicator of developing compressor mechanical problems: impeller rub, labyrinth seal wear, bearing degradation, or aerodynamic instability (surge). A compressor surge event — the condition in which the compressor operating point moves to the left of the surge control line, causing reversed flow through the compressor stages, high-energy oscillatory pressure waves, and potential impeller damage — can destroy the compressor internals in seconds and, in refrigerant service, release the refrigerant inventory to the cold box environment if the compressor casing or discharge piping is breached. AI systems process rendered compressor vibration trend display images — strip chart renders of shaft vibration (mils pp) at the key measuring planes for each compressor body, from the plant process historian — to classify compressor mechanical health and surge proximity: normal, elevated vibration (bearing inspection at next planned outage), high vibration (speed reduction and vibration investigation), and trip threshold (compressor automatic shutdown and MCHE rundown procedure initiation).

An adversarial perturbation on a rendered compressor vibration trend display image that suppresses a rising vibration trend — applying a ±8 DN downward shift to the pixel region encoding the vibration trend line above the elevated alarm level (normalising the apparent vibration to the normal operating band) — causes the compressor monitoring AI to classify a developing bearing failure or surge precursor as normal compressor operation, suppressing the speed reduction and bearing inspection that the elevated vibration classification would initiate. If the undetected vibration trend continues to rise and reaches the automatic trip threshold, the compressor trips and MCHE refrigerant flow is lost: the MCHE begins warming up from −162°C, generating methane boiloff that must be managed through the cold box venting system and flare. If the trip is accompanied by compressor seal failure — a consequence of advanced bearing or labyrinth seal degradation — refrigerant methane is released into the cold box or compressor hall at high pressure and flow rate, creating a significant vapour cloud hazard. NFPA 59A-2023 Section 5.3 requires LNG facilities to have emergency shutdown systems (ESD) that isolate and depressurise process equipment on emergency conditions — adversarial suppression of compressor vibration trend AI eliminates the pre-trip maintenance window during which a controlled compressor shutdown and seal inspection could prevent the escalation to emergency ESD conditions.

4. LNG rundown flow and tank liquid level display AI (Emerson Rosemount LNG flow meter AI, ABB LNG inventory AI, Endress+Hauser Liquiphant tank level AI — LNG rundown and storage inventory AI)

LNG rundown — the continuous flow of liquefied LNG product from the MCHE cold end at −162°C through cryogenic piping to the LNG storage tanks — is the primary product inventory management flow in an LNG liquefaction train. The rundown flow rate and LNG storage tank liquid level are the key parameters for production management and safety: if the LNG storage tank fills beyond the high-level setpoint (typically 95–97% full), excess LNG cannot be accommodated and the MCHE must be shut down or LNG must be recirculated through a pump-back system; if the rundown flow drops below the minimum flow setpoint (due to MCHE underperformance, rundown valve malfunction, or cold box temperature upset), the MCHE warm-end temperature rises and the MCHE begins warming toward the autorefrigeration temperature range where dry-out and hot spots can occur. In LNG storage tanks (full containment tanks with nickel-steel inner containers and prestressed concrete outer containers), an overfill event — where the liquid level rises above the high-level trip setpoint — creates the risk of LNG overflow through the tank roof vents, producing a methane vapour cloud at the tank farm level that, if ignited, produces a flash fire or vapour cloud explosion at the storage tank periphery. AI systems process rendered LNG rundown flow meter displays and storage tank liquid level gauge images to classify LNG inventory status: normal, reduced flow (MCHE performance review required), high tank level (rundown rate reduction and LNG carrier scheduling review required), and high-high tank level (rundown shutdown, MCHE cooldown suspension required).

An adversarial perturbation on a rendered LNG tank level display image that suppresses a rising tank level — applying a ±8 DN downward shift to the pixel region encoding the tank level indicator above the high-level warning setpoint (reducing the apparent tank level from the warning zone to the mid-tank operating range) — causes the inventory management AI to classify an approaching LNG tank overfill as normal mid-range inventory, suppressing the rundown flow reduction and LNG carrier scheduling actions that a high-level classification requires. LNG tank overfill events — such as the Bontang LNG plant tank roof seal failure (Indonesia) and multiple LNG peak-shaving facility overfill incidents documented by FERC in its annual LNG facility inspection reports — are among the most serious LNG facility safety events because the consequences (methane vapour release from tank vents at near-ground level, with ignition potential from the tank farm environment) can produce flash fires or vapour cloud explosions at the facility boundary. PHMSA 49 CFR Part 193.2057 requires LNG storage facilities to have automatic shutoff devices on LNG storage tanks that activate on high-level alarms — adversarial suppression of the tank level AI delays the high-level signal reaching the automatic shutoff system by operating at the AI classification layer above the regulatory compliance boundary.

Integration: LNG liquefaction cold box AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for LNG liquefaction cold box AI belongs at every rendered-image ingestion boundary in the cold box safety AI pipeline — before cold box methane LEL gas detection display AI processes rendered detector display images, before MCHE warm-end temperature profile AI processes rendered thermal distribution images, before refrigerant compressor vibration trend AI processes rendered historian strip chart images, and before LNG rundown flow and tank level AI processes rendered inventory display images. Threshold 30 reflects the Skikda 2004 consequence anchor (27 killed, 3 trains destroyed) and the LNG vapour cloud explosion consequence envelope for undetected cold box methane accumulation.

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"

# LNG liquefaction cold box AI contexts: threshold 30
# NFPA 59A-2023 Standard for LNG Production, Storage and Handling;
# PHMSA 49 CFR Part 193 Liquefied Natural Gas Facilities;
# EN 1473:2016 Installation and Equipment for LNG (European standard).
LNG_COLD_BOX_THRESHOLD = 30


class LNGColdBoxAIContext(Enum):
    COLD_BOX_LEL_GAS    = "cold_box_lel_gas"   # Methane LEL detection display AI
    MCHE_TEMP_PROFILE   = "mche_temp_profile"   # MCHE warm-end temperature profile AI
    COMPRESSOR_VIBRATION = "compressor_vibration" # Refrigerant compressor vibration AI
    LNG_TANK_LEVEL      = "lng_tank_level"       # LNG rundown / tank level inventory AI


class AdversarialLNGColdBoxImageError(Exception):
    """Raised when Glyphward detects adversarial content in an LNG cold box
    AI rendered display image above threshold 30.

    Consequence if not raised:
    - COLD_BOX_LEL_GAS: CH4 accumulation to LEL suppressed → vapour cloud
      ignition in cold box enclosure → VCE → cryogenic piping rupture;
      Skikda 2004 mechanism (27 killed, 3 trains destroyed).
    - MCHE_TEMP_PROFILE: mercury amalgam MCHE degradation suppressed →
      internal gas-refrigerant leakage → cold box methane accumulation.
    - COMPRESSOR_VIBRATION: bearing failure suppressed → compressor seal
      failure → high-pressure refrigerant methane release → VCE hazard.
    - LNG_TANK_LEVEL: tank overfill suppressed → LNG overflow from tank
      roof vents → methane vapour cloud at tank farm → flash fire or VCE.
    Fail-safe: halt AI classification; activate cold box ESD per NFPA 59A
    Section 5.3; require manual gas detector sweep before resuming
    AI-driven cold box safety monitoring classification.
    """

    def __init__(self, scan_id: str, score: int,
                 context: LNGColdBoxAIContext,
                 plant_id: str, train_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.plant_id = plant_id
        self.train_id = train_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial LNG cold box image: "
            f"context={context.value} score={score} "
            f"plant={plant_id} train={train_id} scan_id={scan_id}"
        )


async def scan_lng_cold_box_image(
    image_bytes: bytes,
    context: LNGColdBoxAIContext,
    plant_id: str,
    train_id: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an LNG cold box AI rendered display image for adversarial content.

    Fail-safe contract: AdversarialLNGColdBoxImageError or httpx error →
    halt LNG cold box AI classification for the affected monitoring zone;
    require manual cold box gas detector sweep (NFPA 59A Section 5.3
    gas detection requirements) before resuming AI-driven safety
    monitoring or inventory management classification.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"lng_cold_box:{context.value}:{plant_id}:{train_id}",
        "metadata": {
            "plant_id": plant_id,
            "train_id": train_id,
            "context": context.value,
            "image_sha256": image_hash,
        },
    }
    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"] > LNG_COLD_BOX_THRESHOLD:
        raise AdversarialLNGColdBoxImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            plant_id=plant_id,
            train_id=train_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_lng_cold_box_image at each LNG cold box AI rendered-image ingestion boundary: before cold box methane LEL gas detection display AI (threshold 30), before MCHE warm-end temperature profile AI (threshold 30), before refrigerant compressor vibration trend AI (threshold 30), and before LNG tank level inventory AI (threshold 30). On AdversarialLNGColdBoxImageError for COLD_BOX_LEL_GAS context: immediately initiate cold box ESD sequence per NFPA 59A Section 5.3 — isolate cold box gas detection zone, stop MCHE feed gas flow, activate cold box purge nitrogen, and require manual Photoionisation Detector (PID) sweep by Operations team before restarting. See also: pipeline integrity inspection AI prompt injection (related gas infrastructure AI adversarial context) and urban gas distribution city gate AI prompt injection (related natural gas distribution AI adversarial context). Get early access

Related questions