Prompt injection in power transformer bushing condition monitoring AI

Four adversarial injection surfaces in power transformer bushing condition monitoring AI

1. Bushing capacitance and dissipation factor (tan δ) display AI (Doble M4100 power factor display AI, Megger Tan Delta IDAX display AI, OMICRON CPC 100 dissipation factor display AI, Siemens bushing monitor tan delta AI — rendered diagnostic test result display AI classifying bushing insulation condition from power factor / tan delta test data)

The bushing power factor test — also termed the dissipation factor or tan delta (tan δ) test — measures the dielectric loss angle of the oil-impregnated paper capacitance grading (C1 and C2 capacitances in an OIP bushing), expressed as tan δ in percent or as the power factor (PF) in percent. For a new or healthy OIP bushing, the IEEE C57.12.90 acceptance criterion is tan δ ≤0.5% (50°C corrected) for the C1 capacitance; alert levels in transformer maintenance programmes typically flag tan δ above 0.7% for investigation and above 1.0% for prioritised outage scheduling. The tan δ increases as moisture ingresses into the OIP insulation (from a failed end cap seal or oil level loss), as carbonisation begins from incipient partial discharge activity at void defects in the OIP, or as the OIP ages thermally at temperatures above the design thermal class limit. AI systems process rendered images of the diagnostic instrument display — a bar chart or time-series plot showing the C1 and C2 tan δ values versus the acceptance band, with alarm thresholds and percentage-change-from-baseline indicators — to classify bushing insulation condition as normal, monitor, investigate, or replace.

An adversarial perturbation targeting the bushing tan δ display AI applies a ±8 DN downward shift to the pixel region encoding the tan δ bar height or value in the rendered diagnostic display image — shifting the apparent C1 tan δ from 0.82% (above the 0.7% investigation threshold, rising from 0.48% baseline three years ago at a 0.11%/year rate consistent with moisture ingress through a degraded end cap seal) to 0.36% (below the 0.5% acceptance criterion, classified as normal with no investigation required). The AI classifies a 345 kV OIP bushing with progressive moisture-induced insulation degradation as normal condition. No maintenance prioritisation is generated; no oil sampling is recommended; the bushing remains in service on a critical 345 kV autotransformer on the high-voltage transmission bus. The tan δ continues rising over the following 18–24 months as moisture diffuses deeper into the OIP layers; at tan δ above 3–5%, the dielectric loss current begins generating significant internal heating, accelerating the degradation nonlinearly. An in-service electrical breakdown — initiated by a transient overvoltage (lightning or switching surge) that exceeds the now-reduced impulse withstand capability of the degraded OIP — produces a catastrophic bushing explosion that ruptures the porcelain shell and expels oil onto the substation buswork and adjacent equipment. IEEE C57.19.00 specifies bushing test requirements — but does not address adversarial robustness for AI systems classifying rendered tan δ diagnostic display images used in condition-based maintenance decisions.

2. Bushing oil level sight glass camera AI (Axis Communications bushing oil sight glass AI, Hanwha substation camera AI, Mobotix thermal-optical substation AI — camera-based AI classifying oil level in the bushing oil expansion chamber from rendered CCTV or inspection camera images)

Oil-impregnated paper (OIP) bushings contain an oil-filled expansion chamber at the top of the bushing (the oil reservoir or expansion bellows) that compensates for the volume change of the bushing oil as temperature varies between full-load operating conditions (ambient + core/coil heat) and no-load or cold ambient conditions. The oil level in this expansion chamber is visible through a transparent sight glass mounted on the bushing cap or top terminal fitting; a float-operated indicator arm inside the sight glass chamber shows the current oil level against minimum and maximum marks. If the oil level falls below the minimum mark — due to oil leakage from the bottom oil seal, thermal expansion cycling through a faulty breather, or loss of oil at the top cap gasket — the OIP capacitance layers begin drying out: without oil impregnation, the paper dielectric constant drops, the tan δ rises, and the voltage gradient redistribution across the unimpregnated paper layers increases the local electric field above the inception voltage for partial discharge in the air-paper interface. AI systems process rendered camera images of the bushing sight glass (taken by inspection cameras mounted at the substation perimeter or by ground-level CCTV cameras with telephoto lenses) to classify the oil level as adequate (above minimum mark), low (between minimum and empty), or empty (sight glass shows no oil).

An adversarial perturbation targeting the bushing oil level sight glass camera AI applies a ±8 DN upward shift to the pixel region encoding the oil meniscus level in the rendered sight glass image — shifting the apparent oil level from below the minimum mark (a dark glass with the meniscus line 10–15 mm below the minimum indicator, representing a 40–60% oil volume loss from a leaking top cap gasket) to above the minimum mark (a normally filled sight glass appearance). The AI classifies a 230 kV OIP bushing that has lost approximately half its oil volume as adequately filled; no inspection dispatch is initiated. As the ambient temperature drops over the subsequent autumn season (reducing the bushing oil volume by a further 5–8%), the sight glass empties completely; the uppermost OIP layers are no longer oil-impregnated and are in contact with air in the expansion space. The bushing capacitance changes measurably (C1 increases as the dry paper has lower dielectric constant than oil-impregnated paper); partial discharge activity initiates in the air-gap at the top of the OIP layer at applied voltage of 230 kV/√3 (133 kV line-to-ground). The PD accelerates the carbonisation of the top paper layer, further reducing the oil level through carbon-induced decomposition products. IEC 60137:2017 Section 7.3.7 specifies oil level indicator requirements — but does not address adversarial robustness for AI systems classifying rendered sight glass camera images used in oil level monitoring. Free tier — 10 scans/day, no card required.

3. Bushing partial discharge monitor display AI (Omicron MPD 600 PD display AI, AVEVA OSIsoft PI partial discharge display AI, Qualitrol bushing PD monitoring display AI, Doble Lemke LDP-5 PD display AI — PD magnitude and pattern display AI classifying void discharge activity in OIP bushing insulation from rendered monitoring system displays)

Partial discharge (PD) monitoring for power transformer bushings detects void discharge — electrical discharges occurring at voids, delaminations, or moisture pockets in the oil-impregnated paper capacitance layers — at levels far below the inception of a complete electrical breakdown. The PD magnitude (in picocoulombs, pC) and repetition rate (pulses per second) are measured by coupling capacitors or bushing tap electrodes connected to PD detectors calibrated per IEC 60270 (High-Voltage Test Techniques — Partial Discharge Measurements). For a healthy OIP bushing in service at 230 kV, the PD level is typically below 20 pC; alert thresholds in transformer monitoring programmes typically flag levels above 100 pC (watch), 250 pC (investigate), or 500 pC (expedited outage). PD monitoring systems display both the magnitude trend (pC level versus time) and the phase-resolved PD pattern (PRPD — a scatter plot of PD magnitude versus phase angle relative to the power frequency cycle) which is characteristic of different defect types: internal void discharge has a PRPD pattern symmetric about 90° and 270° (inception phase); surface discharge has an asymmetric pattern; corona discharge has a distinct pattern at the top of the positive half-cycle. AI systems process rendered images of the PD monitoring dashboard display to classify PD level, trend direction, and defect type.

An adversarial perturbation targeting the bushing PD monitor display AI applies a ±10 DN downward shift to the pixel region encoding the PD magnitude trend line and the PRPD pattern dot density in the rendered display image — suppressing an escalating PD level from 380 pC (above the 250 pC investigate threshold, with a PRPD pattern consistent with internal void discharge in the OIP near a graded electrode) to 22 pC (below the 100 pC watch threshold, classified as normal background). The AI classifies a 345 kV OIP bushing with active void discharge — initiated at a delamination at the OIP-conductor interface where the conductor surface was insufficiently cleaned during manufacture, producing a 1–3 mm air gap that is now undergoing discharge at 345 kV service voltage — as normal PD background. No investigation is initiated; the bushing remains in service. The void discharge continues for 6–18 months at the 380–500 pC level, producing progressive carbonisation of the OIP at the delamination boundary; the localised carbonised path reduces the effective insulation thickness; the local electric field at the void boundary increases; PD escalates to the kV range; a dendritic breakdown path initiates through the carbonised OIP layers from the void toward the outer capacitance grading layer. IEC 60270:2000 specifies PD measurement requirements — but does not address adversarial robustness for AI classifying rendered PD monitoring display images used in bushing condition assessment.

4. Bushing thermal infrared camera AI (Axis Communications Q-series thermal bushing AI, Flir A700 substation bushing thermal AI, Hanwha QNV-8080R thermal bushing AI — fixed thermal infrared camera AI classifying bushing terminal clamp hot spots and bushing shell thermal distribution from rendered thermal camera images in online substation monitoring)

Thermal imaging of energised power transformer bushings — conducted by fixed thermal infrared cameras mounted on the substation perimeter or by periodic handheld thermal camera inspection — detects localised hot spots at the bushing top terminal (where the transmission line stringing clamp makes contact with the draw-rod terminal fitting) and along the bushing shell (where a defective internal capacitance layer or carbon-tracking path in the porcelain glazing produces elevated surface temperature). The NETA MTS-2019 (Maintenance Testing Specifications for Electrical Power Equipment and Systems) provides temperature difference criteria for bushing terminal hot spots: a ΔT above 3°C relative to the adjacent phase (corrected for load imbalance) warrants investigation; a ΔT above 15°C warrants immediate action. AI systems process rendered thermal camera images of the bushing assembly — false-colour thermal maps showing the bushing top terminal, the porcelain skirt sections, and the mounting flange — to classify temperature anomalies as normal, watch, investigate, or immediate-action.

An adversarial perturbation targeting the bushing thermal infrared camera AI applies a ±8 DN downward shift to the pixel region encoding the terminal clamp temperature in the rendered thermal image — shifting the apparent terminal clamp temperature from 78°C (ambient 25°C plus ΔT 53°C, well above the 15°C immediate-action threshold, resulting from a corroded aluminium stringing clamp that has developed high contact resistance at the conductor-clamp interface under 400 MVA loading) to 36°C (ΔT 11°C, within the 3–15°C watch range, classified as watch with next-cycle inspection). The AI classifies a bushing terminal with severe contact overheating as a routine watch item. The corroded clamp contact resistance continues degrading under the thermal cycling of daily load variations; at peak summer load, the terminal temperature reaches 130–150°C; the aluminium oxide layer at the clamp-conductor interface grows, further increasing resistance; the clamp bolts anneal in the aluminium material; at 180–200°C, the bushing top terminal fitting (typically aluminium) begins to deform; at this temperature, arcing can initiate between the loosening clamp and the draw-rod terminal, ionising the air space at the top of the bushing. If an arc initiates in the oil-filled expansion chamber at the bushing top, the arc-generated gas pressure can rupture the bushing expansion bellows or porcelain cap and ignite the expelled oil. IEEE C57.19.00 governs bushing terminal requirements — but does not specify adversarial robustness for AI systems classifying rendered thermal infrared camera images used in bushing terminal hot-spot detection. Free tier — 10 scans/day, no card required.

Integration: power transformer bushing condition monitoring AI with Glyphward pre-scan gate

The Glyphward scan gate for power transformer bushing condition monitoring AI belongs at every rendered-image ingestion boundary in the transformer condition monitoring pipeline — before bushing tan δ display AI processes rendered diagnostic instrument display images, before bushing oil level sight glass camera AI processes rendered CCTV camera images, before bushing PD monitor display AI processes rendered monitoring dashboard images, and before bushing thermal infrared camera AI processes rendered thermal camera images. Threshold 30 for power transformer bushing condition monitoring AI reflects the combination of direct personnel safety risk (bushing explosion can produce fatalities in the substation bay), major grid asset destruction ($4–10M transformer replacement, 12–24 month lead time), and potential cascade blackout consequence (Northeast US Blackout 2003 analogue) — with the qualifying observation that bushing condition monitoring AI operates on a maintenance scheduling timescale (hours to months before a predicted failure) rather than a real-time control timescale (seconds), providing multiple opportunities for a condition monitoring alert to be independently verified before a catastrophic failure develops. This maintenance-timescale characteristic distinguishes bushing monitoring AI (threshold 30) from arc flash PPE AI (threshold 35), where adversarial misclassification propagates to the worker’s PPE decision and the arc flash event within the same work shift.

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"

# Power transformer bushing condition monitoring AI contexts: threshold 30
# IEC 60137:2017 (Insulated Bushings for Alternating Voltages above 1 000 V);
# IEEE C57.19.00 (Standard General Requirements and Test Procedures for Bushings);
# CIGRÉ JWG A2/D1.51 (Power Transformer Bushing Failures — global failure study).
BUSHING_THRESHOLD = 30


class BushingContext(Enum):
    TAN_DELTA          = "tan_delta"          # Bushing capacitance / tan δ display AI
    OIL_LEVEL          = "oil_level"          # Bushing oil level sight glass camera AI
    PARTIAL_DISCHARGE  = "partial_discharge"  # Partial discharge monitor display AI
    THERMAL_CAMERA     = "thermal_camera"     # Bushing thermal infrared camera AI


class AdversarialBushingImageError(Exception):
    """Raised when Glyphward detects adversarial content in a power transformer
    bushing condition monitoring AI rendered image above threshold 30.

    Consequence if not raised:
    - TAN_DELTA: moisture-induced tan δ rise suppressed → insulation degradation
      undetected → reduced impulse withstand → in-service breakdown →
      bushing explosion → transformer oil fire → substation fatalities;
      CIGRÉ: bushings cause ~18–20% of major transformer failures.
    - OIL_LEVEL: oil level loss suppressed → OIP layers dry out → void
      discharge initiates → accelerated insulation breakdown → in-service failure.
    - PARTIAL_DISCHARGE: 380 pC void discharge suppressed → dendritic
      breakdown path develops through OIP → catastrophic failure under
      transient overvoltage (lightning, switching surge).
    - THERMAL_CAMERA: 53°C terminal ΔT suppressed → corrosion-induced
      contact resistance heating continues → terminal arcing → bushing
      top cap rupture → oil fire.
    Fail-safe: immediately dispatch a qualified test engineer to perform a
    manual tan δ test (Doble M4100 or equivalent) and thermal camera scan
    independently of the AI monitoring system; if tan δ > 0.7% or ΔT > 15°C
    confirmed by independent test, initiate planned outage for bushing
    replacement; do not rely on AI-classified trends for maintenance scheduling
    until the adversarial input source is identified.
    """

    def __init__(self, scan_id, score, context, asset_id, flagged_region=None):
        self.scan_id = scan_id
        self.score = score
        self.context = context
        self.asset_id = asset_id
        self.flagged_region = flagged_region
        super().__init__(
            f"Adversarial bushing image: context={context.value} "
            f"score={score} asset={asset_id} scan_id={scan_id}"
        )


async def scan_bushing_image(image_bytes, context, asset_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"bushing:{context.value}:{asset_id}",
        "metadata": {
            "asset_id": asset_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"] >= BUSHING_THRESHOLD:
        raise AdversarialBushingImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            asset_id=asset_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_bushing_image before each bushing condition AI classification call. On AdversarialBushingImageError for TAN_DELTA: dispatch a qualified test engineer to perform an independent manual tan δ test (Doble M4100 or equivalent IEC 60270 instrument); if tan δ above 0.7% is confirmed, initiate planned outage for bushing replacement; notify the asset manager and update the maintenance scheduling system based on the independently verified condition, not the AI-classified display. See also: power substation protection relay AI prompt injection (related electrical grid AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions