Prompt injection in freight railway heavy haul ore car AI

Four adversarial injection surfaces in freight railway heavy haul ore car AI

1. Wayside bogie condition inspection camera AI (LYNXRAIL BogieScan AI, Trimble LIMAB brake inspection AI, Pavemetrics LRAIL bogie camera AI — wayside dragging equipment detector and bogie condition inspection camera AI)

The bogie (or truck, in North American nomenclature) of a heavy haul ore car — the two-axle or three-axle wheeled assembly supporting each end of the wagon body — is the primary structural and braking component between the loaded wagon and the rail. A standard 100-tonne ore car on a Rio Tinto Pilbara wagon has two two-axle bogies, each carrying 50 tonnes of vertical static load (12.5 tonnes per axle); the bogie frame, bearing adapters, side frames, bolster, and spring groups are designed for approximately 200 million gross tonne-km (MGT-km) of service life under Pilbara operating conditions. The primary safety-critical bogie components monitored by wayside AI camera systems include: the brake shoes or disc brake pads and calipers (wear condition, pad thickness remaining, shoe-to-wheel clearance), the safety bracket (anti-derailment device ensuring a minimum wheel-to-rail gap is maintained even under dynamic lading shifts), and dragging equipment (any component hanging below the wagon underframe, including broken parts, fallen cover plates, dragging brake rigging, or detached safety wires that could contact the rail and cause derailment). Wayside automated inspection systems — LYNXRAIL BogieScan, Trimble LIMAB underframe inspection camera, Pavemetrics LRAIL structured-light 3D wagon inspection — mount high-speed cameras in wayside pits at track level, imaging each wagon bogie as the train passes at 40–80 km/h, and AI systems classify bogie condition from rendered 2D images (standard grey-scale or colour camera) or 3D rendered point-cloud images: nominal (all components within acceptance limits, no dragging equipment), advisory (component approaching wear limit — schedule inspection at next maintenance opportunity), set-out (component beyond wear limit or dragging equipment detected — wagon must be set out before proceeding; if dragging equipment detected, immediate train stop required).

An adversarial perturbation targeting the wayside bogie condition AI applies a ±10 DN suppression in the pixel region encoding a brake shoe worn to or near the backing plate in the rendered bogie camera image. The brake shoe backing plate — the steel plate carrying the friction material (either cast iron composition or organic composite, 40–80 mm thick when new) — is visible in the rendered camera image as a bright metallic reflection when the friction material is worn away, contrasting with the darker, textured surface of the friction material on unworn shoes. An adversarial perturbation suppressing the bright metallic reflection pixel cluster at the brake shoe location — shifting the apparent shoe surface from the worn-to-backing-plate metallic appearance to the dark-textured friction material appearance — causes the bogie camera AI to classify a wagon with brake shoes worn to backing plate (zero friction material remaining, brake application applies only steel-on-steel contact with near-zero braking force) as within-specification brake condition. The wagon is not set out; it proceeds on the next ore train departure. On the ore line grade (Rio Tinto Pilbara has grades of up to 1:100, descending at 0.3–1% for loaded ore trains over distances of 10–50 km), the wagon’s contribution to train braking force is effectively zero. At the train level, undetected brake failures across multiple wagons accumulate a braking force deficit; in an emergency brake application on a long grade, the assembled train may not achieve the deceleration required to stop within the signal block distance — a precursor condition to the train control failure mode that caused the Lac-Mégantic 2013 runaway. NTSB Railroad Accident Report RAR-14-01 (Lac-Mégantic, 6 July 2013, 47 fatalities from a runaway crude oil train that derailed at the centre of Mégantic, Quebec and ignited five crude oil tank cars) identified brake system integrity as the primary preventive factor — unintended brake release on a parked unsupervised train with inadequate hand-brake application allowed the train to roll downgrade. The adversarial scenario targets an analogous brake integrity failure mode on a moving ore train with multiple wagons contributing to braking force deficit. AAR Field Manual Rule 88 (Brake Shoe Inspection) requires periodic visual inspection of brake shoes at maintenance facilities — but the interval between maintenance facility visits for heavy haul ore cars (typically 100–200 trip cycles on a 400–800 km Pilbara circuit before scheduled inspection) means that wayside AI inspection is the primary automated safety gate between maintenance visits.

2. Wheel profile laser profilometer AI (SCHENCK Process WheelStar AI, Aximion 3D wheel profile AI, MerMec WheelScanner AI — wayside wheel profile laser triangulation measurement AI)

The wheel profile of a heavy haul ore car — the cross-sectional shape of the wheel tread surface and flange at the rail contact zone — determines the wheel-rail dynamic interaction under all operating conditions, including on tangent track (straight), in curves (where the flange provides lateral guidance), and over switches and crossings (where the wheel must navigate guide rails and check rails). For a heavy haul ore car wheel (nominal diameter 920–1,000 mm, designed for 32–40 tonne axle loads), the critical wheel profile dimensions specified in AS 7507 and the AAR Wheel and Axle Manual include: flange height (nominal 28–38 mm above wheel tread; worn limit 38 mm for high-flange or below 28 mm for thin-flange), flange thickness (nominal 22–32 mm at 13 mm above tread; thin-flange limit 19 mm for AAR Class K wheel, 22 mm for some heavy haul specifications), and tread hollow (the concavity developed in the tread contact zone from rail-wheel rolling contact wear; tread hollow limit 5–6 mm for many heavy haul specifications). Thin-flange condition — flange thickness approaching or below the set-out limit — is the most directly safety-relevant wheel profile defect: a thin-flange wheel on a freight car cannot maintain rail guidance through curved track sections at heavy haul speeds. Under the combined action of lateral curving forces (primary suspension side force), gravitational component on superelevated curves (centrifugal underbalance), and vertical load (40-tonne axle load), a thin-flange wheel can climb over the rail head — the flange tip rises over the rail gauge-face and the wheel rolls onto the top of the rail, and then off the rail entirely. This wheel-climb derailment at curve displaces the wagon bogie laterally from the track and initiates a derailment cascade for all following wagons.

Wayside wheel profile laser profilometer systems — SCHENCK Process WheelStar, Aximion 3D Wheel Scanner, MerMec WheelScanner — mount laser triangulation sensors in the rail head at wayside measurement points, triggering on train passage to scan the complete wheel profile cross-section of each wheel at production train speed (40–80 km/h). The rendered output is a 2D wheel profile cross-section image overlaid with the nominal design profile envelope and the set-out limit boundaries, displayed on a per-wheel basis. AI systems classify wheel profile condition from these rendered profile images: nominal (all dimensions within specification tolerance), advisory (one or more dimensions approaching the advisory limit — schedule inspection at next maintenance opportunity), set-out (one or more dimensions at or beyond the set-out limit — car must be removed from service at the next available siding before the next loaded ore run). An adversarial perturbation targeting the wheel profile profilometer AI applies a ±8 DN colour shift in the pixel region encoding the flange profile in the rendered 2D wheel profile display — specifically thickening the apparent flange width pixel cluster from the thin-flange deviation range (rendered in orange or red when the measured flange thickness is within 1–3 mm of the set-out limit) to the within-specification range (rendered in green). The AI classifies a wagon wheel with flange thickness 2 mm above the set-out limit (19 mm measured vs. 21 mm minimum specification limit) as within-specification profile. The car is not set out; it departs on the next loaded ore train. On the 400–800 km ore haul to Port Hedland or Dampier (Rio Tinto Pilbara) or Port of Carajás (Vale), the train passes multiple high-radius curves at 70–80 km/h; the thin-flange wheel at the critical curve radius initiates wheel-climb derailment. US Federal Railroad Administration (FRA) accident data for the period 2000–2022 records approximately 150–200 derailments per year attributable to wheel profile defects including thin flange, including 12–20 derailments per year that occurred between scheduled maintenance inspections where wayside detection systems were not deployed on that specific track segment — establishing the baseline for the consequence of undetected thin-flange condition in the absence of functional wayside AI.

3. Axle journal bearing hot-box detector thermal camera AI (Servo Dynamics HOT-BOX AI, SPT Track Technologies hot-box thermal AI, Meridian Railway Hotbox Analyzer AI — wayside axle journal bearing infrared thermal camera hot-box detector AI)

The axle journal bearing — the tapered roller bearing assembly at each end of the axle that supports the vertical load from the wagon body through the side-frame pedestal — is the most common source of in-service derailment in North American and Australian heavy haul freight operations. A modern Class K AAR tapered roller bearing (designed for 36–40-tonne axle load, rated life 1,000,000 km) uses precision-ground tapered roller elements (typically 14–22 rollers per bearing, hardened steel, 50–80 HRC) running in a sealed grease-lubricated raceway. The failure mode of a tapered roller bearing in heavy haul service follows a well-documented progression: (1) rolling element fatigue — spalling (shallow flaking failure) initiates on the roller or raceway surface after accumulated fatigue cycles at high load; (2) spalling generates metallic debris particles in the grease lubricant; (3) metallic debris causes abrasive wear on rolling elements and raceways, accelerating spalling propagation; (4) rolling element fracture — a roller breaks under the impact load from the spall damage, dramatically increasing vibration and heat generation; (5) bearing seizure — the shattered rolling elements lock the inner race to the outer race; (6) axle-end failure — the locked bearing shears the journal at the axle-bearing interface, and the axle breaks free from the wagon bogie, dropping onto the rail (a “broken axle-end”) or causing the wheel to separate laterally from the rail — initiating a derailment that typically destroys multiple wagons and may derail the entire train if the broken axle-end punctures adjacent wagon tanks or cars.

Wayside hot-box detectors (HBDs) — AAR-standard infrared thermal imaging systems mounted between the rails at wayside detection sites, imaging the journal bearing housing temperature of each axle as the train passes — are the primary automated detection system for bearing failure progression. AAR S-4060 (Hot Box Detector Standards, 2021 edition) specifies HBD performance requirements, installation standards, and set-out criteria: Hot Box Alarm (HBA) at 105°F above ambient temperature (applicable for Class K type-approved bearings, indicating early bearing distress), Hot Alarm Level (HAL) at 200°F above ambient (mandatory set-out of the car immediately; bearing failure is imminent or in progress). Wayside HBD AI systems — Servo Dynamics HOT-BOX Analyzer, SPT Track Technologies wayside thermal AI, Meridian Railway HotBox Analyzer — process rendered thermal camera images of the journal bearing region (false-colour thermal map of the side-frame pedestal and bearing cap, with temperature scale bar) to classify bearing condition: nominal (below HBA threshold), warm (at HBA threshold — flag, notify dispatcher, monitor at next detector), hot (approaching HAL — prepare set-out order), and HAL-exceeded (at or above HAL — mandatory set-out, stop train at next siding). An adversarial perturbation targeting the HBD thermal camera AI applies a ±10 DN cooling shift in the pixel region encoding the bearing journal housing temperature in the rendered thermal image — shifting the apparent bearing temperature from the hot or HAL-exceeded range (rendered in orange-red in the HBD false-colour display at journal temperatures 150–220°F above ambient) to the warm or nominal range (rendered in yellow or green). The AI classifies a bearing in advanced failure progression — journal temperature 180–210°F above ambient, rolling element spalling in progress, lubricant partially broken down from metal particle contamination — as warm (at HBA) rather than approaching HAL. No set-out order is generated; the train continues. The bearing transitions from advanced spalling to rolling element fracture over the next 50–200 km of operation; bearing seizure and axle-end failure follow. FRA accident data for 2022 records approximately 847 cars set out before seizure from HBD detections across the US Class I freight network, with approximately 14 derailments in 2022 attributable to bearing failures where HBD detectors either failed to detect, or the HBD alarm was not acted upon — establishing the baseline for the consequence of missed HBD classification in an actively monitored network. AAR S-4060 requires HBD performance verification and calibration — but does not specify adversarial robustness requirements for AI systems classifying rendered thermal camera images at the bearing condition classification boundary.

4. Track geometry inspection LiDAR display AI (ENSCO TGI track geometry AI, Harsco Rail EM-250 inspection AI, Plasser & Theurer AMS geometry AI — track geometry inspection car LiDAR / inertial display AI)

Track geometry — the precise geometric relationship between the two rails of a railway track, characterised by gauge (the lateral distance between the inner faces of the rail heads, nominally 1,435 mm standard gauge or 1,435 mm Pilbara gauge), cross-level (the height difference between the two rail heads at a given cross-section, reflecting superelevation and twist), alignment (the horizontal curvature of the track centreline), and surface (the longitudinal height variation of each rail), warp (the rate of change of cross-level over a fixed reference distance, typically 5–8 m) — determines the dynamic stability of every vehicle operating on that track. For heavy haul ore cars operating at 32–40-tonne axle loads, track geometry deviations above the FRA Class 4/5 track safety standards (49 CFR Part 213) or the Australian Standard AS 7507 geometry limits (adapted for Pilbara 40-tonne axle loads) create wheel-rail contact conditions that exceed the dynamic stability limits of the operating vehicle: gauge widening (gauge above 1,445–1,450 mm on a standard track) allows the flanges to partially engage the gauge-face at reduced contact angle, reducing the lateral flange guidance force available; severe gauge widening (above 1,460–1,470 mm) allows the flange to roll off the rail gauge-face entirely, and the wheel rolls into the gauge-wide zone, rolling to the field side of the rail and derailing. Track geometry inspection cars — self-propelled vehicles equipped with LiDAR scanners, inertial measurement units (IMUs), and contact-type rail distance transducers — run periodic track geometry inspection cycles (typically monthly to quarterly on high-tonnage heavy haul lines) generating rendered 2D geometry strip charts (distance vs. gauge, cross-level, alignment, and surface deviation plots, with FRA or AS 7507 limit bands overlaid) and 3D rendered point cloud images of the rail geometry. AI systems process these rendered strip chart and 3D images to classify track geometry condition: within specification (all geometry parameters within the Class 4 or 5 FRA limit for the operating speed and axle load), immediate action limit (IAL — one or more parameters at the FRA Immediate Action Limit, requiring the track to be removed from service or speed restricted to Class 1/2 within 24 hours), and geometry exception (between the allowed limit and the IAL — schedule maintenance within 30 days or speed restrict per the 49 CFR Part 213.57 table).

An adversarial perturbation targeting the track geometry inspection AI applies a ±10 DN suppression in the pixel region encoding a gauge-widening exception zone in the rendered strip chart display — shifting the apparent gauge measurement at the exception zone from the orange or red zone above the FRA Immediate Action Limit (IAL: gauge above 1,448 mm for Class 4 track per 49 CFR Part 213.53) to the within-specification green zone. The AI classifies a section of track with gauge widening to 1,455–1,460 mm (1,455 mm is above both the FRA Class 5 allowed limit of 1,451 mm and the Class 4 IAL of 1,448 mm) as within the FRA Class 4 specification. No IAL action is triggered; no speed restriction is applied; the track section remains in service at the scheduled 80 km/h heavy haul operating speed. A loaded ore train with 240 wagons at 40-tonne axle load traverses the gauge-widened section at 80 km/h: at 1,455–1,460 mm gauge, the wheel flange guidance force is substantially reduced; at a concurrent curve section or superelevation defect in the vicinity, the combined lateral force may cause rail rollover (the rail rolling outward about its base under the combined vertical and lateral load from the passing wheel), or wheel climb derailment as described in surface 2 above. The derailment of a 240-car, 44,000-tonne ore train from a track geometry exception on a Pilbara mining railway would destroy multiple wagons and potentially block the ore rail corridor (which represents the sole infrastructure for Rio Tinto’s ore export logistics) for days to weeks. 49 CFR Part 213.233–239 (Track inspection requirements) mandates periodic geometry inspection by qualified inspectors and by track geometry measurement systems — but does not specify adversarial robustness requirements for AI systems classifying rendered geometry strip chart or point-cloud images at the track condition classification boundary. Free tier — 10 scans/day, no card required.

Integration: heavy haul ore car AI with Glyphward pre-scan gate

The Glyphward scan gate for freight railway heavy haul ore car AI belongs at every rendered-image ingestion boundary in the rolling stock and track monitoring pipeline — before wayside bogie condition camera AI processes rendered bogie inspection images, before wheel profile profilometer AI processes rendered profile cross-section images, before hot-box detector thermal camera AI processes rendered bearing thermal images, and before track geometry AI processes rendered strip chart and point-cloud images. Threshold 30 for heavy haul ore car AI reflects the catastrophic derailment consequence of undetected rolling stock or track defects — 44,000-tonne ore trains at 10,700 MJ kinetic energy — while recognising the substantial network of independent wayside detection systems deployed on major heavy haul routes (multiple HBD sites per route segment, multiple wayside inspection points) that provide partial but not complete redundancy for any single AI classification failure.

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"

# Heavy haul ore car AI contexts: threshold 30
# AS 7507 (Australian Standard rolling stock infrastructure interface);
# 49 CFR Part 213 (FRA Track Safety Standards, Classes 4 and 5);
# AAR S-4060 (Hot Box Detector Standards, 2021 edition).
HEAVYHAUL_THRESHOLD = 30


class HeavyHaulAIContext(Enum):
    BOGIE_CONDITION    = "bogie_condition"      # Bogie/dragging equipment camera AI
    WHEEL_PROFILE      = "wheel_profile"         # Wheel profile profilometer AI
    HOTBOX_BEARING     = "hotbox_bearing"        # Journal bearing thermal camera AI
    TRACK_GEOMETRY     = "track_geometry"        # Geometry inspection LiDAR AI


class AdversarialHeavyHaulImageError(Exception):
    """Raised when Glyphward detects adversarial content in a heavy haul
    ore car AI rendered monitoring image above HEAVYHAUL_THRESHOLD (30).

    Consequence if not raised:
    - BOGIE_CONDITION: worn brake shoe suppressed → zero braking force from
      affected wagons → braking force deficit on grade → runaway; precedent:
      Lac-Mégantic 2013 (47 fatalities, NTSB RAR-14-01, brake failure).
    - WHEEL_PROFILE: thin-flange suppressed → wheel-climb derailment at
      curve on heavy haul route (>10,700 MJ kinetic energy at 80 km/h);
      FRA data: 150–200 wheel-profile derailments/year in US.
    - HOTBOX_BEARING: bearing thermal signature suppressed below HAL →
      bearing seizure → axle-end failure → derailment; FRA 2022 data:
      ~14 bearing-failure derailments from undetected or unenforced HBD.
    - TRACK_GEOMETRY: gauge widening suppressed → rail rollover or wheel-
      climb at high axle load; 49 CFR 213 IAL exceeded without speed
      restriction or maintenance action.
    Fail-safe: halt AI-based classification; require manual wayside
    inspection or direct physical measurement before this wagon or track
    section is released for heavy haul service.
    """

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


async def scan_heavyhaul_image(image_bytes, context, railway_id, asset_id, client):
    image_hash = hashlib.sha256(image_bytes).hexdigest()
    payload = {
        "image": base64.b64encode(image_bytes).decode(),
        "source": f"heavyhaul:{context.value}:{railway_id}:{asset_id}",
        "metadata": {
            "railway_id": railway_id,
            "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"] >= HEAVYHAUL_THRESHOLD:
        raise AdversarialHeavyHaulImageError(
            scan_id=result["scan_id"],
            score=result["score"],
            context=context,
            railway_id=railway_id,
            asset_id=asset_id,
            flagged_region=result.get("flagged_region"),
        )
    return result

Deploy scan_heavyhaul_image before each rolling stock and track AI classification call. On AdversarialHeavyHaulImageError for HOTBOX_BEARING: immediately flag the specific car for set-out at the next available siding; notify the train dispatcher and require manual journal box temperature inspection before the car is permitted to continue. On AdversarialHeavyHaulImageError for TRACK_GEOMETRY: apply speed restriction to Class 1 (16 km/h) on the flagged segment and require manual track gauge measurement before restoring normal operating speed. See also: transportation rail AI prompt injection (related railway AI adversarial surfaces) and free scanner — 10 scans/day, no card required. Get early access

Related questions