Prompt injection in offshore wind farm inspection AI

Four adversarial injection surfaces in offshore wind farm inspection AI

1. Blade leading edge erosion drone photogrammetry AI (Clobotics SCOPE, Sulzer Blade Repair AI, LM Wind Power GE Vernova AI)

Blade leading edge erosion drone inspection uses autonomous or semi-autonomous drones carrying calibrated RGB cameras to capture complete 360° blade surface photogrammetry during turbine idling or at low rotor speed. The drone imagery is processed into orthomosaic images — geometrically corrected plan-view renders of the full blade surface area at 0.5–2 mm/pixel ground sampling distance — that AI classification pipelines segment and classify into erosion severity zones across the blade radius. Clobotics SCOPE Wind AI (deployed at Equinor’s Hywind Tampen and Ørsted’s Hornsea farm assets), Sulzer Electra blade repair AI, and GE Vernova blade inspection drone AI use convolutional segmentation models trained on DNV GL blade inspection grading reference images to classify each surface pixel into Stage 1–5 erosion categories and generate a blade condition report that drives repair prioritisation — Stage 3–5 regions require repair campaign scheduling; Stage 4–5 regions require immediate decommissioning of the affected blade if the spar cap is exposed or cracked.

An adversarial perturbation on a rendered blade orthomosaic image that downgrades a Stage 4 erosion zone — shifting the exposed fibre-reinforcement beige/tan colour of a delaminated composite ply toward the intact gel-coat grey of surrounding undamaged blade surface by a ±12 DN pixel shift in the affected orthomosaic patch — causes the blade AI to classify the region as Stage 2 (surface paint degradation, no immediate repair required) rather than Stage 4 (composite delamination, repair within 3 months per O&M contract SLA). A Stage 4 delamination left unrepaired for an offshore wind maintenance cycle — typically 6–12 months between planned major maintenance access windows — can propagate to Stage 5 spar cap crack and eventually produce an uncontrolled blade release. A 55-tonne offshore wind blade (at Vestas V236-15 MW or Siemens Gamesa SG 14-236 DD scale) released from a 120m+ hub height produces a kinetic energy impact event with a safety exclusion radius of 500m — extending to the inter-turbine cable route where subsea cable maintenance vessels may be operating.

2. Monopile scour detection sonar bathymetry AI (Fugro Seawatch AI, SBM Offshore SIEM AI, Bureau Veritas Digital Twin AI)

Monopile foundations — the large-diameter (6–11 m) steel cylinder piles driven 30–40 m into the seabed that support over 90% of installed offshore wind turbines — are susceptible to local scour: the erosion of seabed sediment around the monopile circumference caused by wave and current-induced bed shear stress. Scour depth is defined as the vertical distance from the undisturbed seabed level to the lowest point of the scour hole around the monopile, measured by multibeam echo sounder (MBES) or single-beam sonar bathymetry surveys. Fugro Seawatch marine survey AI, SBM Offshore SIEM (Structured Integrity and Erosion Monitoring) AI, and Bureau Veritas offshore wind digital twin AI process rendered MBES bathymetry images — false-colour depth maps of the seabed around each monopile, with colour scale encoding water depth (blue: deepest scour; green/yellow: undisturbed seabed) at sub-metre resolution — to classify scour depth relative to the design allowable scour (S/D ratio, where S = scour depth and D = monopile diameter). DNV-ST-0126 (“Support Structures for Wind Turbines”) Section 4.3.3 requires that scour protection (rock armouring or concrete mattresses) maintain S/D ≤ 1.3 for monopiles; S/D exceedances increase the effective unsupported pile length, reducing the monopile’s lateral resistance and potentially causing resonant fatigue loading under wave excitation at the turbine’s natural frequency.

An adversarial perturbation on a rendered MBES bathymetry image that fills a scour hole — shifting the blue deep-scour colour signature of a 3.5m scour hole around a 7m-diameter monopile (S/D = 0.5, approaching the 1.3 limit) to the green undisturbed-seabed colour of the surrounding bed by a hue rotation within the false-colour scale — causes the scour AI to report S/D < 0.1 (scour protection intact, no action required) rather than S/D = 0.5 (within limits but approaching threshold, monitor monthly). A scour hole that continues to develop undetected beyond S/D = 1.3 without triggering the rock placement remediation campaign specified by the turbine’s foundation design (per DNV-ST-0126 Section 4.3) reduces the monopile’s lateral stiffness and raises the turbine’s first natural frequency, potentially bringing it within the rotor’s 1P excitation frequency range — a resonance condition that produces accelerated fatigue damage in the monopile weld zones and tower base, potentially exceeding the foundation’s 25–35 year fatigue design life within 5–10 years.

3. J-tube cable bend stiffener inspection AI (Oceaneering ROV AI, Saab Seaeye Falcon ROV AI, IRM AI offshore wind)

Inter-array cables connecting offshore wind turbines to the substation export cable are routed through J-tubes — curved steel tubes fixed to the monopile transition piece that guide the subsea cable from the seabed horizontal lay to the vertical entry into the turbine foundation — with polyurethane bend stiffeners fitted at the J-tube exit point to limit the cable’s bending radius under wave and current loading. Bend stiffener cracking or disbondment from the cable sheath allows seawater ingress into the cable’s metallic armour wires, causing galvanic corrosion, eventual conductor insulation failure, and inter-array cable electrical fault requiring expensive cable repair or replacement (estimated £500k–£2M per cable repair including diving or ROV vessel mobilisation). Oceaneering International’s ROV inspection AI (deployed at Ørsted, SSE Renewables, and ScottishPower Renewables offshore wind farms), Saab Seaeye Falcon ROV inspection AI, and generic IRM (Inspection, Repair and Maintenance) ROV camera frame AI process underwater camera images of J-tube entries and bend stiffeners — captured during annual or biennial ROV inspection campaigns — to identify bend stiffener cracking, splitting, chalking degradation, or disbondment from the cable sheath jacket.

An adversarial perturbation on an ROV camera frame image of a J-tube bend stiffener that suppresses a lateral crack signature — smoothing the crack’s dark shadow line against the polyurethane’s grey surface by blending the shadow tone into surrounding material texture within ±10 DN — causes the ROV inspection AI to classify the bend stiffener as “intact, no defects identified,” missing a Stage 2 crack (surface crack penetrating less than 25% of the stiffener cross-section) that would normally trigger a “monitor at next scheduled inspection” recommendation. A missed Stage 2 crack that propagates to full stiffener fracture before the next scheduled ROV campaign allows unrestrained cable bending at the J-tube exit, accelerating cable fatigue at the bend zone and, in severe North Sea metocean conditions (Hs > 4 m wave height, currents > 1.5 m/s), producing cable conductor fatigue fracture within 2–3 years — resulting in an inter-array cable electrical fault that disconnects all turbines on the cable string from the offshore substation, producing an unplanned loss of 20–50 MW of generating capacity.

4. Tower external corrosion drone thermal AI (Drone Volt Hercules AI, Percepto Autonomous AI, Applanix POSPac WV AI)

Offshore wind turbine tower sections — fabricated from rolled steel plate in 20–30m cylindrical cans — are susceptible to external coating breakdown and metal corrosion in the tidal splash zone (between mean low water and mean high water) and in the atmospheric zone above the splash zone, where salt spray and high humidity drive accelerated coating degradation on the inner and outer tower surfaces. Drone Volt Hercules autonomous inspection drone AI, Percepto autonomous inspection AI (deployed at multiple Equinor and BP offshore facilities), and photogrammetric tower inspection AI from Applanix/PCI Geomatics process RGB and thermal infrared drone inspection images of tower external surfaces — at 2–5 mm/pixel resolution on the atmospheric zone and at lower resolution for the splash zone where wave action limits drone approach — to classify coating condition (intact, chalking, blistering, delaminated) and identify areas of active metal corrosion (rust staining, pitting, steel plate thickness loss) that require coating repair or protective re-application. Thermal infrared drone imagery of tower surfaces identifies hot-spots associated with water ingress behind delaminated coating — a precursor to accelerated under-coating corrosion that is not visible in RGB imagery until the coating system fails.

An adversarial perturbation on a rendered drone RGB tower inspection orthomosaic that shifts the rust-staining orange-brown signature of active corrosion on a coastal tower joint — tinting the rust colour toward the grey of intact coating within ±8 DN in the orthomosaic render — causes the tower corrosion AI to classify the joint as “coating intact, no corrosion identified,” suppressing an area of active corrosion that requires immediate coating repair before metal section loss reaches the Class 3 corrosion threshold requiring plate thickness inspection under DNV-ST-0126 Section 6.4.3. In the offshore wind splash zone — continuously wetted by wave action and salt spray — un-coated steel corrodes at 0.3–0.5 mm/year average rate; a suppressed active corrosion zone of 0.5 m² in the splash zone of a 100 MW offshore wind tower can, over a 12-month inspection cycle, progress to a 0.5 mm section loss in the tower shell — approaching the corrosion allowance specified in the original tower design (typically 0.3–1.0 mm depending on coating system design life and maintenance strategy).

Integration: offshore wind farm inspection AI scanning with Glyphward pre-scan gate

The Glyphward scan gate for offshore wind farm inspection AI belongs at each rendered-image ingestion boundary — before blade erosion drone orthomosaic AI processes photogrammetry images, before scour detection AI processes MBES bathymetry renders, before J-tube cable inspection AI processes ROV camera frames, and before tower corrosion AI processes drone RGB and thermal images. Threshold 40 for offshore wind inspection AI reflects the availability of alternative inspection access methods (rope access technicians for blades, divers for foundation inspection) that provide complementary verification before structural decisions are made.

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"

# Offshore wind farm inspection AI contexts: threshold 40
# IEC 61400-3-1 (Offshore wind turbines, 2019 ed.); DNV-ST-0126.
WIND_FARM_AI_THRESHOLD = 40


class WindFarmAIContext(Enum):
    BLADE_EROSION_ORTHO   = "blade_erosion_ortho"   # Blade LEP drone photogrammetry AI
    MONOPILE_SCOUR_SONAR  = "monopile_scour_sonar"  # MBES bathymetry scour depth AI
    JTUBE_ROV_CAMERA      = "jtube_rov_camera"      # J-tube bend stiffener ROV AI
    TOWER_CORROSION_DRONE = "tower_corrosion_drone" # Tower RGB/IR drone corrosion AI


class AdversarialWindFarmImageError(Exception):
    """Raised when Glyphward detects adversarial pixel content in an offshore
    wind farm inspection AI rendered image above threshold 40.

    Consequence if not raised: blade Stage 4 delamination misclassified
    as Stage 2 → repair deferred → Stage 5 spar crack → 35-55 t uncontrolled
    blade release at 120m hub height; or monopile S/D exceedance undetected →
    resonant fatigue loading → foundation weld fatigue failure.
    Fail-safe: defer AI inspection classification; escalate to rope-access
    or ROV manual re-inspection per DNV-ST-0126 Section 4.3 requirements.
    """

    def __init__(self, scan_id: str, score: int,
                 context: WindFarmAIContext,
                 wind_farm_id: str, turbine_id: str,
                 flagged_region: dict | None = None) -> None:
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.wind_farm_id = wind_farm_id
        self.turbine_id = turbine_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial wind farm image: "
            f"context={context.value} score={score} "
            f"farm={wind_farm_id} turbine={turbine_id} scan_id={scan_id}"
        )


async def scan_wind_farm_image(
    image_bytes: bytes,
    context: WindFarmAIContext,
    wind_farm_id: str,
    turbine_id: str,
    blade_position: str | None,
    inspection_date: str,
    client: httpx.AsyncClient,
) -> dict:
    """Scan an offshore wind farm inspection AI image for adversarial content.

    Fail-safe contract: AdversarialWindFarmImageError or httpx error →
    defer AI inspection classification; escalate to manual re-inspection per
    DNV-ST-0126 Section 4.3 inspection programme requirements. For blade AI:
    require rope-access technician visual confirmation of eroded zones before
    issuing repair deferred classification.

    Args:
        image_bytes: Blade orthomosaic, MBES bathymetry render, ROV camera
            frame, or drone RGB/thermal tower image bytes.
        context: WindFarmAIContext identifying the inspection modality.
        wind_farm_id: Wind farm identifier (e.g., "Hornsea 2", "Vineyard Wind 1").
        turbine_id: Turbine ID within the farm (e.g., "OWF-T042").
        blade_position: Blade identifier — "B1", "B2", "B3" or None.
        inspection_date: ISO 8601 date of the inspection campaign.
        client: Shared httpx.AsyncClient for connection reuse.

    Returns:
        Glyphward scan result dict.

    Raises:
        AdversarialWindFarmImageError: if score exceeds threshold 40.
        httpx.HTTPStatusError: on Glyphward API error (fail-closed).
    """
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"wind:{context.value}:{wind_farm_id}:{turbine_id}",
        "metadata": {
            "wind_farm_id": wind_farm_id,
            "turbine_id": turbine_id,
            "blade_position": blade_position,
            "inspection_date": inspection_date,
            "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()

    await _write_wind_farm_scan_audit(
        image_hash=image_hash,
        scan_id=result["scan_id"],
        score=result["score"],
        context=context,
        wind_farm_id=wind_farm_id,
        turbine_id=turbine_id,
        blade_position=blade_position,
        inspection_date=inspection_date,
        flagged=result["score"] > WIND_FARM_AI_THRESHOLD,
    )

    if result["score"] > WIND_FARM_AI_THRESHOLD:
        raise AdversarialWindFarmImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            wind_farm_id=wind_farm_id,
            turbine_id=turbine_id,
            flagged_region=result.get("flagged_region"),
        )
    return result


async def _write_wind_farm_scan_audit(
    *, image_hash: str, scan_id: str, score: int,
    context: WindFarmAIContext, wind_farm_id: str, turbine_id: str,
    blade_position: str | None, inspection_date: str, flagged: bool,
) -> None:
    record = {
        "ts": datetime.now(timezone.utc).isoformat(),
        "scan_id": scan_id,
        "image_sha256": image_hash,
        "context": context.value,
        "score": score,
        "threshold": WIND_FARM_AI_THRESHOLD,
        "flagged": flagged,
        "wind_farm_id": wind_farm_id,
        "turbine_id": turbine_id,
        "blade_position": blade_position,
        "inspection_date": inspection_date,
        "regulatory_refs": [
            "IEC 61400-3-1:2019 (Offshore wind turbines — Design requirements)",
            "DNV-ST-0126 (Support Structures for Wind Turbines, 2021)",
            "DNV-ST-0376 (Rotor Blades for Wind Turbines, 2015)",
            "DNV-GL Blade Inspection Guideline (Rev. 0, 2019)",
            "IEC 61400-1 Ed.4:2019 (Wind turbines — Design requirements)",
            "BIMCO/GDV Offshore Wind Farm Safety Guidelines",
            "UK MCGA MCA Safety Management Code for Offshore Renewable Energy Installations",
        ],
    }
    audit_path = Path("/var/log/glyphward/wind_farm_ai_scan_audit.jsonl")
    audit_path.parent.mkdir(parents=True, exist_ok=True)
    with audit_path.open("a") as fh:
        fh.write(json.dumps(record) + "\n")

Deploy scan_wind_farm_image at each offshore wind inspection AI rendered-image boundary: before blade erosion orthomosaic AI (threshold 40), before monopile scour sonar AI (threshold 40), before J-tube ROV camera AI (threshold 40), and before tower corrosion drone AI (threshold 40). On AdversarialWindFarmImageError: defer AI inspection classification; escalate to rope-access technician visual inspection (for blades) or diver/ROV manual re-inspection (for foundation and J-tube) per DNV-ST-0126 Section 4.3 inspection programme requirements. For blade AI: do not issue Stage 2 or lower classification based on adversarially flagged orthomosaic images without rope-access visual confirmation. Get early access

Related questions

Further reading

• Prompt injection in subsea and offshore inspection AI — ROV camera AI bypass, FPSO hull inspection adversarial attacks
• Prompt injection in transmission line inspection AI — NERC CIP gap, drone RGB+thermal+UV inspection adversarial attacks
• Prompt injection in renewable energy AI — solar panel defect detection AI adversarial attacks
• Prompt injection in dam safety monitoring AI — InSAR deformation AI bypass, seepage detection adversarial attacks
• Prompt injection scanning API free tier — 10 scans/day, no card required