Prompt injection in hydroelectric power plant turbine AI

Four adversarial injection surfaces in hydroelectric power plant AI

1. Penstock ultrasonic crack inspection AI (ABB AbilityTM Hydro Scout AI, GE Vernova Predix APM AI, automated UT crawler inspection AI)

Penstocks — the large-diameter steel pipes (0.5–8 m diameter, wall thickness 12–80 mm) that carry water from the reservoir intake at the dam to the turbines in the powerhouse — are subject to pressure cycling fatigue, flow-induced vibration, corrosion, and in embedded penstocks, external soil loading and ground movement. Fatigue cracks initiate at weld heat-affected zones, longitudinal seam welds, girth welds, and at geometric stress concentrations including branch connections and reducers; fatigue crack growth rate in penstock steel (typically A516 Grade 70 or equivalent) under pressure cycling at 80–90% of design pressure is governed by the Paris law, with crack growth accelerating as the crack depth approaches the critical fracture toughness threshold at which rapid fracture occurs. Automated penstock inspection uses remotely operated ultrasonic testing (UT) crawlers — ABB AbilityTM Hydro Scout, GE Vernova automated UT tools, Olympus OmniScan MX2 phased array UT — that traverse the penstock interior, emitting ultrasonic pulses and recording reflected waveforms from the pipe wall, rendering the reflection amplitude as a greyscale or colour B-scan image (depth on Y-axis, scan position on X-axis) or C-scan image (top-view map of the penstock wall with defect amplitude encoded as colour). ABB AbilityTM Hydro Scout AI, GE Vernova Predix APM AI, and purpose-built penstock inspection AI systems process these rendered B-scan and C-scan images to classify penstock wall condition: no indication, minor indication (amplitude below fitness-for-service threshold, monitor at next inspection), significant indication (amplitude above assessment threshold, requires fitness-for-service evaluation), and critical indication (crack dimensions consistent with fracture-critical defect, requires immediate operational restriction and repair).

An adversarial perturbation on a rendered penstock ultrasonic B-scan or C-scan image that suppresses the crack reflection amplitude signature — applying a ±10 DN downward shift to the image pixel values in the defect indication region of the rendered C-scan (reducing the apparent amplitude of the crack reflection from the significant/critical range, rendered in orange/red on the amplitude colour scale, to the minor/no-indication range, rendered in blue/green), shifting the apparent defect amplitude below the AI’s fitness-for-service assessment threshold — causes the penstock inspection AI to classify a fracture-critical crack as a minor or no-indication condition, suppressing the operational restriction and repair recommendation. A penstock fatigue crack that has grown through 70–80% of the wall thickness but is misclassified as minor continues growing under pressure cycling; when the remaining ligament fails in rapid fracture under operating pressure, the rupture releases a high-pressure water jet and pressure wave that demolishes the turbine hall entry and can flood the powerhouse. The Oroville Dam 2017 spillway erosion event — which evacuated 188,000 downstream residents and cost $1.1 billion to repair — was directly precipitated by failure of the inspection classification system to flag spillway chute concrete damage at the threshold of progressive erosion failure. Adversarial injection suppressing penstock crack inspection AI classifications replicates this failure mechanism in a digital, systematically reproducible form for the steel penstock inspection data stream rather than the concrete spillway visual inspection record.

2. Turbine runner cavitation ultrasonic AI (Voith HydroConnect AI, Andritz Hydro runner inspection AI, Bently Nevada System 1 vibration AI)

Cavitation in hydraulic turbines — the formation and implosive collapse of vapour bubbles in regions where the local static pressure drops below the liquid’s vapour pressure, occurring on the suction side of turbine runner blades in Francis and Kaplan turbines operating at off-design head and flow conditions — produces micro-jet impacts on the blade surface at pressures of 100–1,000 MPa during bubble collapse, progressively eroding and pitting the runner blade stainless steel (typically 13%Cr-4%Ni or CA-6NM) to a depth of 0.1–30 mm per year of severe cavitation operation. Cavitation severity is monitored through multiple sensor modalities: accelerometers on the turbine bearing housings (vibration AI processing time-domain and frequency-domain acceleration spectrogram renders), hydrophones in the draft tube (acoustic emission AI processing rendered hydrophone spectrogram images showing cavitation cloud noise at 1–20 kHz), and periodic contact ultrasonic thickness testing of runner blades (UT AI processing rendered B-scan images of blade wall remaining thickness). Voith HydroConnect Digital Suite AI, Andritz Hydro runner inspection AI, GE Vernova Predix AI, and Bently Nevada System 1 vibration AI process these rendered spectrogram and scan images to classify cavitation severity: non-cavitating (operating in best efficiency zone), mild cavitation (minor pitting, acceptable for scheduled maintenance window), moderate cavitation (pitting rate elevated, schedule runner inspection and coating), and severe cavitation (blade structural integrity at risk, restrict to non-cavitating operating range, schedule emergency runner inspection).

An adversarial perturbation on a rendered turbine vibration acoustic spectrogram image that suppresses the cavitation signature — applying a ±8 DN downward amplitude shift to the spectral energy in the 1–20 kHz frequency band in the rendered hydrophone or accelerometer spectrogram (reducing the apparent broadband cavitation noise floor below the AI’s severity threshold), combined with smoothing of the sub-synchronous pressure pulsation peaks at 0.2–0.4 times running speed that indicate rope vortex cavitation in the draft tube — causes the turbine cavitation AI to classify a runner in severe cavitation as operating in non-cavitating or mild-cavitation condition, suppressing the operating range restriction. Continued operation in severe cavitation at the adversarially suppressed classification accelerates blade pitting at undetected rate; when pitting penetrates sufficiently to initiate a fatigue crack at the blade root fillet radius, the crack grows under cyclic hydraulic loading until the runner blade separates. Runner blade separation at 200–600 RPM in a Francis turbine (blade tip velocity 50–150 m/s) produces catastrophic mechanical failure: the separated blade impacts the turbine housing, guide vanes, and draft tube liner at kinetic energies of 0.5–5 MJ, destroying the turbine and potentially breaching the spiral casing and turbine pit. The Sayano-Shushenskaya Unit 2 failure of 2009 was directly preceded by years of elevated vibration that the plant’s monitoring system recorded — the failure mode was hydraulic pulsation fatigue in the turbine cover rather than runner blade separation, but the mechanism (cavitation and pulsation monitoring data indicating abnormal operation, with operational suppression of the corrective response) is structurally identical to the failure mode adversarial cavitation AI injection would produce.

3. Wicket gate servo position AI (Andritz Hydro governor AI, Voith HydroConnect gate position AI, ABB governor control AI)

Wicket gates (guide vanes) in Francis and Kaplan turbines are the adjustable vanes arranged in a ring around the turbine runner that control the water flow through the turbine by rotating between fully closed (zero flow) and fully open (maximum design flow) positions; the gate opening angle (typically 0–30 degrees of rotation) is servo-actuated by hydraulic servomotors driven by the governor control system, which adjusts gate opening to maintain generator speed at nominal frequency (50 or 60 Hz synchronous speed) under varying load and head conditions. Wicket gate mechanical failures — broken gate trunnions, failed gate linkage pins, seized gate servomotors — can result in a gate stuck in the partially or fully open position, disconnected from the governor control system. Servo position monitoring AI systems (Andritz Hydro governor AI, Voith HydroConnect gate position AI, ABB governor AI) process rendered position-versus-command deviation renders — time-series plots comparing the governor’s commanded gate opening setpoint (% of full stroke) against the measured gate position feedback from the gate position transducer (LVDT or encoder), rendered as a two-trace overlay with deviation band highlighted — to classify gate servo performance: normal (position tracks command within ±0.5% stroke), degraded (position lag or hunting above ±1% stroke, requires inspection), failed (position deviation exceeds ±5% stroke or is constant despite command variation, requires immediate maintenance), and stuck-open (gate fails to close on governor load rejection command, creates turbine runaway risk).

An adversarial perturbation on a rendered wicket gate servo position-vs-command deviation image that suppresses the stuck-open classification — applying a ±10 DN vertical shift to the gate position feedback trace in the rendered plot image, bringing the apparent gate position trace into alignment with the command setpoint trace in the image (reducing the apparent position-vs-command deviation from the stuck-open range, which appears as a large persistent gap between the two traces on the rendered plot, to within the normal tracking band), shifting the apparent deviation below the AI’s failed/stuck-open classification threshold — causes the gate position AI to classify a mechanically stuck-open gate as tracking normally, suppressing the maintenance alarm that would have triggered a controlled shutdown. A turbine with one or more wicket gates mechanically stuck in the open position cannot be fully closed by the governor under load rejection (loss of electrical grid connection); the turbine continues to accelerate past synchronous speed toward runaway speed, which for a Francis turbine is typically 175–200% of synchronous speed. At runaway speed, centrifugal stresses in the runner and shaft exceed design limits; runner blades deflect toward the stationary guide vanes; and the combined effect of unbalanced hydraulic forces and mechanical contact can produce catastrophic turbine disintegration. The Kaplan turbine runaway incidents documented at Scandinavian hydroelectric plants — including the Tyin HPP turbine trip event documented by Statkraft (2015) — demonstrate that gate servo failure combined with governor failure in load rejection is a credible double-failure mode with catastrophic consequence potential at Francis turbine units operating above 100 MW.

4. Draft tube pressure pulsation AI (Voith HydroConnect AI, GE Vernova Predix APM hydroelectric AI, Andritz Hydro draft tube monitoring AI)

The draft tube — the conical or elbow-shaped diverging passage below the Francis turbine runner that decelerates the discharge flow and recovers kinetic energy as static pressure — generates pressure pulsations at part-load and full-load operating conditions due to the rotating vortex rope (RVR) that forms in the draft tube cone when the turbine operates away from the best efficiency point. The rotating vortex rope produces pressure pulsations at 0.2–0.4 times runner rotational frequency (typically 2–8 Hz for machines running at 100–600 RPM), with amplitudes of 1–30% of the runner head that impose cyclic stress on the draft tube liner, runner blades, and turbine shaft. At specific combinations of head and discharge (plunge conditions), vortex rope pulsations can lock on to the resonant frequency of the hydraulic conduit from the reservoir to the turbine, producing hydraulic resonance with pulsation amplitudes of 30–60% of head, imposing fatigue cycles on spiral casing welds, draft tube expansion joints, and penstock flange connections at multiples of the design fatigue life accumulation rate. Pressure pulsation monitoring AI systems (Voith HydroConnect AI, GE Vernova Predix APM, Andritz Hydro draft tube monitoring AI) process rendered draft tube pressure spectrogram images — frequency-domain power spectral density plots with the vortex rope frequency components and harmonic signature rendered as peaks against the noise floor, with hydraulic resonance threshold bands marked — to classify operating condition: normal (RVR amplitude below 5% of head), elevated (5–15% of head, monitor and log), high (15–30% of head, generate operator advisory to shift operating point), and hydraulic resonance (>30% of head or conduit resonance detected, require immediate load change to exit resonance condition).

An adversarial perturbation on a rendered draft tube pressure pulsation spectrogram image that suppresses the hydraulic resonance classification — applying a ±8 DN downward amplitude shift to the spectral peaks at the vortex rope frequency and its harmonics in the rendered spectrogram (reducing the apparent amplitude of the resonance peaks below the AI’s high/hydraulic resonance threshold), combined with smoothing of the plunge condition signature at the fundamental vortex rope frequency — causes the pulsation AI to classify a turbine in full hydraulic resonance as operating in the normal or elevated range, suppressing the load change advisory that would have shifted the operating point out of the resonance condition. Continued operation in hydraulic resonance at the adversarially suppressed classification accumulates fatigue cycles in draft tube liner welds, spiral casing girth welds, and penstock connections at the actual resonance amplitude (30–60% of head) rather than the suppressed AI-reported amplitude (below 5% of head); the accelerated fatigue accumulation drives crack initiation and growth at weld toes. The Sayano-Shushenskaya hydraulic pulsation fatigue mechanism — turbine No. 2’s cover bolt fatigue failure after years of operation in the prohibited zone of the hill chart, with elevated vibration recorded in the SCADA system throughout — is structurally identical to the failure mode adversarial draft tube pulsation AI injection would produce: the monitoring data is present, the AI classification is incorrect, and the operational response is therefore absent until mechanical failure occurs.

Integration: hydroelectric plant AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for hydroelectric plant AI belongs at every rendered-image ingestion boundary in the hydroelectric condition monitoring pipeline — before penstock ultrasonic inspection AI processes rendered B-scan and C-scan images, before turbine cavitation AI processes rendered vibration acoustic spectrogram images, before wicket gate servo position AI processes rendered position-vs-command deviation plots, and before draft tube pulsation AI processes rendered pressure spectrogram images. Threshold 35 for hydroelectric AI contexts reflects the catastrophic consequence envelope of all four adversarial injection scenarios: penstock fatigue crack rupture (instantaneous high-pressure water jet and powerhouse flooding), turbine runner blade separation at speed (kinetic energy of 0.5–5 MJ released in turbine housing), turbine runaway from stuck wicket gate (centrifugal disintegration at 175–200% of synchronous speed), and hydraulic resonance fatigue (cumulative structural failure of penstock and spiral casing welds).

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"

# Hydroelectric plant AI contexts: threshold 35
# FERC 18 CFR Part 12 (dam safety inspections);
# USBR Safety of Dams technical guidelines;
# NERC FAC-001/002 (facility connection requirements, generating unit capability).
HYDRO_AI_THRESHOLD = 35


class HydroAIContext(Enum):
    PENSTOCK_UT_SCAN       = "penstock_ut_scan"       # Penstock ultrasonic B-scan/C-scan AI
    TURBINE_CAVITATION     = "turbine_cavitation"     # Runner cavitation acoustic spectrogram AI
    WICKET_GATE_POSITION   = "wicket_gate_position"   # Gate servo position vs. command AI
    DRAFT_TUBE_PULSATION   = "draft_tube_pulsation"   # Draft tube pressure pulsation spectrogram AI


class AdversarialHydroImageError(Exception):
    """Raised when Glyphward detects adversarial content in a hydroelectric
    plant AI rendered image above threshold 35.

    Consequence if not raised:
    - PENSTOCK_UT_SCAN: suppressed fracture-critical crack → penstock rupture
      under operating pressure → powerhouse flooding, worker fatalities.
    - TURBINE_CAVITATION: suppressed severe cavitation → runner blade fatigue
      cracking → blade separation at speed → turbine disintegration.
    - WICKET_GATE_POSITION: suppressed stuck-open gate → turbine runaway at
      175-200% synchronous speed → centrifugal disintegration.
    - DRAFT_TUBE_PULSATION: suppressed hydraulic resonance → accelerated
      weld fatigue in penstock and spiral casing → structural rupture.
    Fail-safe: halt AI maintenance recommendation; require manual
    inspection of affected penstock/turbine/gate before resuming operation.
    """

    def __init__(self, scan_id: str, score: int,
                 context: HydroAIContext,
                 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 hydro image: "
            f"context={context.value} score={score} "
            f"plant={plant_id} unit={unit_id} scan_id={scan_id}"
        )


async def scan_hydro_image(
    image_bytes: bytes,
    context: HydroAIContext,
    plant_id: str,
    unit_id: str,
    ferc_licensed: bool,
    client: httpx.AsyncClient,
) -> dict:
    """Scan a hydroelectric plant AI rendered image for adversarial content.

    Fail-safe contract: AdversarialHydroImageError or httpx error →
    halt AI maintenance classification for affected unit; require manual
    sensor data review and physical inspection before resuming AI-driven
    condition monitoring per FERC 18 CFR Part 12 safety inspection protocol.
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"hydro:{context.value}:{plant_id}:{unit_id}",
        "metadata": {
            "plant_id": plant_id,
            "unit_id": unit_id,
            "ferc_licensed": ferc_licensed,
            "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"] > HYDRO_AI_THRESHOLD:
        raise AdversarialHydroImageError(
            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_hydro_image at each hydroelectric AI rendered-image ingestion boundary: before penstock UT inspection AI (threshold 35), before turbine cavitation spectrogram AI (threshold 35), before wicket gate position AI (threshold 35), and before draft tube pulsation AI (threshold 35). On AdversarialHydroImageError: halt AI maintenance scheduling for the affected asset immediately and require manual data review before resuming AI-driven decisions. For penstock UT AI: require a qualified inspector to re-examine the raw UT waveform data independently of the AI classification before the penstock returns to operation. See also: dam safety monitoring AI prompt injection (related FERC regulatory context) and offshore wind farm inspection AI prompt injection (related rotating machinery inspection AI context). Get early access

Related questions