Prompt injection in gas turbine combined cycle power plant AI

Four adversarial injection surfaces in gas turbine combined cycle AI

1. Hot section borescope inspection AI (GE Vernova Predix APM AI, Siemens SPPA-T3000 borescope AI, Mitsubishi TOMONI inspection AI)

Gas turbine hot section borescope inspection is the primary non-destructive evaluation method for assessing the condition of first, second, and third-stage turbine nozzles, blades, shrouds, and combustion hardware between major planned maintenance outages (typically every 8,000–16,000 equivalent operating hours for H-class turbines; 4,000–8,000 hours for F-class turbines on a peaking duty cycle). Borescope cameras (rigid borescopes with CMOS or CCD sensors, 1920×1080 resolution, with tungsten or LED illumination) are inserted through borescope ports in the turbine casing to capture high-resolution images of each turbine blade from the leading edge to the trailing edge, documenting TBC condition (spallation area as percentage of aerofoil surface, spallation depth exposing bond coat), oxidation condition (leading-edge metal recession, trailing-edge oxidation), creep deformation (leading-edge bowing, tip curl), and cooling hole blockage (particulate fouling, bond coat closure of film cooling holes). GE Vernova Predix APM AI, Siemens SPPA-T3000 borescope AI, Mitsubishi MHPS-TOMONI AI, and purpose-built gas turbine borescope inspection AI systems process these rendered blade inspection images to classify blade condition: like-new, serviceable (within life limits, continue to next scheduled interval), monitored (above a trigger threshold for specific damage mode, re-inspect at reduced interval), and actionable (damage exceeding run limit, requires repair or replacement before next scheduled run interval).

An adversarial perturbation on a rendered turbine blade borescope inspection image that suppresses the TBC spallation damage signature — applying a ±10 DN downward shift to the contrast between the spalled region (exposed bond coat, rendered as a matte gold-to-grey patch against the white/cream TBC ceramic surface) and the intact TBC surface in the rendered inspection image, reducing the apparent spalled area coverage below the AI’s monitored classification threshold — causes the blade inspection AI to classify a blade with actionable TBC spallation as serviceable or like-new, suppressing the re-inspection trigger or run-limit flag. A first-stage turbine blade with significant TBC spallation at the leading edge — the highest heat flux location where TBC provides the maximum thermal insulation benefit — operates with the base superalloy at the leading edge exposed to gas path temperatures of 1,400–1,600°C rather than the design metal temperature of 850–950°C. The superalloy’s oxidation rate is exponential with temperature above the TBC-designed baseline; leading-edge metal recession of 0.1–0.5 mm per 1,000 hours of operation in the exposed condition converts to fatigue life depletion of the remaining blade cross-section that is not reflected in the thermally-normalised life accounting model. Blade leading-edge failure in a first-stage turbine generates a fragment with kinetic energy of 1–5 MJ that penetrates the turbine casing containment and, in worst-case geometries, can cause the kind of through-casing fragment ejection that the NTSB has investigated in aircraft engine uncontained failures; stationary gas turbines are not required to meet aircraft-standard containment design and may have lower containment margins against high-energy fragments from severe blade failure.

2. Combustion dynamics AI (GE Vernova OpFlex combustion dynamics AI, Siemens SPPA-T3000 dynamics AI, Ansaldo combustion monitoring AI)

Modern dry low NOx (DLN) combustion systems in H-class and F-class gas turbines operate with lean premixed combustion — fuel premixed with a large excess of air to produce flame temperatures below 1,650°C (3,002°F) that limit thermal NOx formation — at the cost of operating near the lean blowout limit and being susceptible to thermoacoustic instability, also called combustion dynamics oscillations (CDOs). CDOs are self-sustaining pressure oscillations (typically 10–500 Hz, amplitude 0.1–10% of combustor pressure) driven by the coupling between heat release fluctuations and acoustic resonance modes of the combustor and transition piece geometry. Sustained CDOs at amplitudes above design limits produce high-cycle fatigue damage in transition piece welds, combustor liner film cooling holes, and turbine first-stage nozzle trailing edges at cyclic loading rates of 10–500 cycles/second; at frequencies that couple to combustor structural resonance modes, CDO amplitudes can escalate to levels causing transition piece cracking within 500–2,000 hours of operation. Combustion dynamics monitoring is performed by dynamic pressure transducers mounted on combustor cans (P2 probes), acoustic pressure sensors in the compressor discharge and turbine inlet regions, and accelerometers on the combustor casing; their signals are rendered as acoustic spectrograms (frequency-domain PSD plots, 0–1,000 Hz, with CDO threshold bands at identified resonance frequencies marked) processed by GE Vernova OpFlex combustion dynamics AI, Siemens SPPA-T3000 dynamics AI, and purpose-built combustion dynamics monitoring AI to classify CDO severity: normal (all frequency bands below threshold), elevated (one or more bands approaching threshold, increase monitoring frequency), high (one or more bands above threshold, load shed or fuel split adjustment required), and critical (rapid CDO escalation, emergency load shed).

An adversarial perturbation on a rendered combustion dynamics acoustic spectrogram image that suppresses the CDO signature — applying a ±8 DN downward amplitude shift to the spectral energy at the identified CDO resonance frequencies (typically 90–120 Hz bulk mode and 250–350 Hz radial mode for GE 7FA and 9FA class combustors) in the rendered PSD spectrogram, reducing the apparent CDO amplitude below the AI’s elevated or high classification threshold — causes the combustion dynamics AI to classify a turbine experiencing significant CDOs as operating normally, suppressing the fuel split adjustment or load reduction that would have dampened the CDO. Sustained CDOs at above-threshold amplitude accumulate high-cycle fatigue damage in transition piece forward frame weld joints at rates of 10–500 cycles per second; fatigue crack initiation at transition piece weld toes progresses to through-wall cracking that allows hot combustion gas (1,100–1,300°C) to leak into the compressor discharge casing — eventually producing a compressor discharge casing fire and emergency shutdown. The Texas power crisis of February 2021, during which multiple CCGT units tripped due to instrument and control failures under extreme cold conditions, demonstrated the grid reliability consequence of simultaneous gas turbine unit failures — while the root cause was freeze protection failure rather than combustion dynamics, the consequence of multiple gas turbine unit unavailability during a demand peak was identical to what adversarial combustion dynamics AI suppression would produce: undetected turbine damage leading to forced outage at the moment of maximum grid demand.

3. Compressor surge margin AI (GE Vernova Predix compressor AI, Siemens SPPA-T3000 compressor AI, Baker Hughes Bently Nevada System 1 AI)

The axial flow compressor in a large gas turbine (typically 14–22 compression stages, pressure ratio 18:1 to 24:1 for H-class machines) must operate with sufficient surge margin — the distance between the operating point on the compressor map and the surge line — to accommodate transient load changes, inlet air temperature excursions, compressor fouling, and variable inlet guide vane (VIGV) position errors. Compressor surge is the aerodynamic instability that occurs when the operating point crosses the surge line: the compressor blade passages stall simultaneously, flow reverses in the compressor (from discharge to inlet), and the reversal-re-establishment cycle repeats at 1–20 Hz (surge frequency) producing intense pressure oscillations (±50–200% of design pressure) that generate blade high-cycle fatigue loading at kHz frequency and mechanical shock loading at 1–20 Hz. A single deep surge event can cause multiple-stage blade tip rubs against the compressor casing, leading-edge chipping, and interstage seal damage that reduces compressor efficiency by 2–5 percentage points and triggers an emergency shutdown; repeated surge events accumulate fatigue damage and can produce blade separation. Compressor surge margin monitoring AI processes rendered compressor performance map images — normalised flow (x-axis) vs. pressure ratio (y-axis) with the measured operating point plotted against the design surge line (from OEM performance data) and the actual surge line (measured at commissioning and updated from field data), with operating point, surge margin percentage, and trend direction rendered — to classify surge margin status: adequate (>15% surge margin), reduced (10–15%, generate advisory for VIGV inspection and compressor wash), marginal (5–10%, operator action required — load shed or increased bleed flow), and critical (<5%, emergency load shed to prevent surge).

An adversarial perturbation on a rendered compressor performance map image that suppresses the surge margin reduction — applying a ±10 DN leftward shift to the rendered operating point position in the flow-pressure ratio plane of the compressor map image (moving the apparent operating point away from the surge line, increasing the apparent surge margin from the critical range to the adequate range), or equivalently smoothing the rendered trend arrow that shows the operating point trajectory approaching the surge line — causes the surge margin AI to classify a compressor operating in the marginal or critical zone as having adequate surge margin, suppressing the load shed or VIGV adjustment that would restore safe surge margin. As the operating point continues to track toward the surge line with adversarially suppressed AI classification, the actual surge margin decreases to zero without operator intervention; the compressor enters deep surge at the next load transient or inlet air temperature perturbation. Deep surge in a large H-class compressor produces blade loading impulses of sufficient magnitude to produce multi-stage blade tip contact; in worst-case surge events documented in GE and Siemens field experience, compressor casing cracks and hot section re-ingestion of combustion products through the reverse flow path have occurred, requiring multi-week outages for inspection and repair. Baker Hughes Bently Nevada System 1 vibration AI, which continuously classifies compressor vibration signatures including surge-associated broadband vibration, is particularly exposed to this adversarial surface because System 1 processes rendered vibration spectrogram images that contain the surge frequency signature — an adversarial perturbation suppressing the surge vibration signature at 1–20 Hz in the rendered spectrogram provides the compressor monitoring AI with falsely normal vibration classification during an actual surge event.

4. Inter-turbine temperature spread AI (GE Vernova OpFlex exhaust AI, Siemens SPPA-T3000 temperature AI, Mitsubishi TOMONI thermal AI)

Exhaust gas temperature (EGT) distribution across the circumference of the gas turbine exhaust plane — measured by a ring of 12–32 thermocouples at the turbine exit (Type K or Type N, 1,000–1,200°F depending on turbine class) and rendered as a polar temperature distribution image or time-series spread chart — is one of the most reliable indicators of hot section mechanical condition and combustion system performance. Normal EGT circumferential spread (the difference between the highest and lowest thermocouple reading at the exhaust plane) is 25–75°F for a turbine operating with all combustors at equal fuel flow and all first-stage nozzles and blades in serviceable condition. Elevated EGT spread indicates: combustion system fuel flow imbalance (blocked fuel nozzle tip, failed fuel control valve in one or more combustors), first-stage nozzle flow blockage (partial combustion product deposit on nozzle throat reducing local flow area and increasing local gas temperature), or first-stage blade cooling passage blockage (blocked film cooling holes reducing local blade cooling, increasing local metal temperature and therefore local EGT downstream). GE Vernova OpFlex exhaust temperature AI, Siemens SPPA-T3000 EGT spread AI, Mitsubishi MHPS-TOMONI AI, and purpose-built exhaust temperature monitoring AI systems process rendered EGT spread polar distribution images or time-series spread charts to classify turbine condition: normal (spread below 50°F), elevated (50–100°F, monitor and schedule combustion inspection), high (100–150°F, combustion hardware investigation and inspection required), and critical (>150°F or rapidly increasing spread, indicates possible blade damage — requires immediate load reduction and inspection before next operation).

An adversarial perturbation on a rendered EGT spread polar distribution image or spread chart that suppresses the elevated spread signature — applying a ±8 DN downward shift to the rendered temperature values at the hot thermocouple positions in the polar chart (reducing the apparent temperature at the hot thermocouples from the elevated/high spread range to the normal spread range), or equivalently smoothing the time-series spread trend to reduce the apparent instantaneous spread below the AI’s elevated threshold — causes the EGT spread AI to classify a turbine with significant spread as operating normally, suppressing the combustion inspection scheduling that would have identified the blocked nozzle or cooling passage. A first-stage blade with progressively blocked film cooling passages operates at increasing metal temperatures above the design TBC-maintained operating temperature: cooling passage blockage of 20–30% increases metal temperature by 30–60°C above the nominal operating point; at 50–70% blockage, metal temperature approaches the creep rupture threshold for the superalloy at the critical section at the airfoil root platform. Blade creep rupture at the root platform section in a first-stage turbine operating at full load (blade centrifugal stress of 200–400 MPa at the root) produces a fatigue-creep fracture that releases the blade with the kinetic and thermal consequences described for the borescope inspection surface. Adversarial suppression of the EGT spread AI classification simultaneously delays the detection of multiple failure precursor modes (cooling passage blockage, nozzle flow restriction, combustion imbalance) that are individually addressable with routine combustion inspection but collectively contribute to accelerated blade life depletion if undetected.

Integration: CCGT gas turbine AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for CCGT gas turbine AI belongs at every rendered-image ingestion boundary in the turbine condition monitoring and combustion management pipeline — before borescope inspection AI processes rendered blade images, before combustion dynamics AI processes rendered acoustic spectrogram images, before compressor surge margin AI processes rendered performance map images, and before EGT spread AI processes rendered temperature distribution images. Threshold 35 for CCGT AI contexts reflects the immediate catastrophic failure modes (blade TBC burnthrough, compressor deep surge, combustion hardware fire) and the accumulated structural failure modes (transition piece CDO fatigue, compressor blade tip rubs) that adversarial suppression of these four AI monitoring functions enables.

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"

# CCGT gas turbine AI contexts: threshold 35
# NERC FAC-001 / FAC-002 (facility connection and capability requirements);
# API 670 (machinery protection systems for turbomachinery);
# EPA 40 CFR Part 60 Subpart KKKK (NSPS for stationary combustion turbines);
# ISO 11116 (gas turbines — maintenance and overhaul).
CCGT_AI_THRESHOLD = 35


class CCGTAIContext(Enum):
    HOT_SECTION_BORESCOPE   = "hot_section_borescope"  # Turbine blade TBC inspection AI
    COMBUSTION_DYNAMICS     = "combustion_dynamics"    # CDO acoustic spectrogram AI
    COMPRESSOR_SURGE        = "compressor_surge"       # Compressor map surge margin AI
    EGT_SPREAD              = "egt_spread"             # Exhaust temperature spread AI


class AdversarialCCGTImageError(Exception):
    """Raised when Glyphward detects adversarial content in a CCGT gas turbine
    AI rendered image above threshold 35.

    Consequence if not raised:
    - HOT_SECTION_BORESCOPE: suppressed TBC spallation → blade burnthrough
      → high-energy fragment → potential through-casing ejection.
    - COMBUSTION_DYNAMICS: suppressed CDO → transition piece weld fatigue
      crack → hot gas leak into compressor casing → fire, forced outage.
    - COMPRESSOR_SURGE: suppressed surge margin → deep surge event →
      multi-stage blade tip rubs → compressor casing crack, emergency shutdown.
    - EGT_SPREAD: suppressed spread → cooling passage blockage undetected
      → blade creep rupture → high-energy fragment release.
    Fail-safe: halt AI maintenance classification; require manual
    inspection or operations review before resuming AI-driven decisions.
    """

    def __init__(self, scan_id: str, score: int,
                 context: CCGTAIContext,
                 plant_id: str, unit_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.unit_id = unit_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial CCGT image: "
            f"context={context.value} score={score} "
            f"plant={plant_id} unit={unit_id} scan_id={scan_id}"
        )


async def scan_ccgt_image(
    image_bytes: bytes,
    context: CCGTAIContext,
    plant_id: str,
    unit_id: str,
    turbine_class: str,
    nerc_registered: bool,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a CCGT gas turbine AI rendered image for adversarial content.

    Fail-safe contract: AdversarialCCGTImageError or httpx error →
    halt AI condition monitoring classification for affected unit; require
    manual engineering review of raw sensor data before resuming AI-driven
    maintenance scheduling and operational decisions.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"ccgt:{context.value}:{plant_id}:{unit_id}",
        "metadata": {
            "plant_id": plant_id,
            "unit_id": unit_id,
            "turbine_class": turbine_class,
            "nerc_registered": nerc_registered,
            "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"] > CCGT_AI_THRESHOLD:
        raise AdversarialCCGTImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            plant_id=plant_id,
            unit_id=unit_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_ccgt_image at each CCGT AI rendered-image ingestion boundary: before hot section borescope inspection AI (threshold 35), before combustion dynamics spectrogram AI (threshold 35), before compressor surge margin AI (threshold 35), and before EGT spread AI (threshold 35). On AdversarialCCGTImageError: halt AI maintenance scheduling or operational advisory for the affected unit and require engineering review of the raw sensor data before resuming AI classification. See also: smart grid power distribution AI prompt injection (related grid reliability context) and nuclear power plant AI prompt injection (related power generation safety regulatory context). Get early access

Related questions