Arc flash incident energy AI adversarial injection: how ±8 DN in the rendered PPE category display misclassifies a Category 3 arc flash hazard as Category 2 — and why NFPA 70E-2021 has no adversarial robustness criterion for the IEEE 1584-2018 arc flash analysis AI

The arc flash energy breakopen threshold: why PPE category misclassification is a fatal failure mode

Arc-rated PPE — arc flash suits, face shields, balaclava hoods, arc-rated gloves, and arc-rated switching coats — is designed and rated under ASTM F1959/F1959M (Standard Test Method for Determining the Arc Rating of Materials for Clothing) to provide two related protections: prevention of onset second-degree burns through intact fabric (the ATPV criterion), and prevention of breakopen that would expose the skin to direct thermal contact with arc plasma or ignited underlayers (the EBT criterion). Both values — ATPV and EBT — are expressed in cal/cm² and represent the incident energy level at which there is a 50% probability of the specified failure occurring in an arc flash event of that energy level. The arc rating of a garment or system is the lower of its ATPV and EBT under NFPA 70E definitions.

The critical distinction between these two failure modes for the adversarial injection threat model is their physical mechanism. ATPV failure — onset second-degree burns through intact arc-rated fabric — means the fabric did not break open but transmitted sufficient thermal energy to begin the process of dermal damage at the Stoll threshold (the skin temperature/time relationship that defines onset second-degree burn). The worker survives; the PPE did its structural job but transmitted more energy than it was designed to absorb at that rating level. EBT failure is categorically different: the fabric develops holes under the thermal load, exposing the fabric layers beneath — typically cotton or cotton-blend base layers, since NFPA 70E Section 130.7(C)(11) requires that clothing worn under arc flash PPE be of natural fibres (cotton, wool) rather than synthetic fibres that would melt. Cotton fabric ignites at an incident energy of approximately 4–5 cal/cm² — well below the EBT of even Category 2 PPE (8 cal/cm²). At the EBT failure point, the residual incident energy at the hole location above the EBT value — 10.4 cal/cm² in a Category 3 event where Category 2 PPE has been donned — ignites the cotton underlayer in direct thermal contact with the skin. The resulting contact burns produce full-thickness third-degree dermal destruction in less than 100 milliseconds.

The PPE category structure in NFPA 70E Table 130.7(C)(15)(a) is engineered to maintain a safety margin above the EBT of each rated layer across the incident energy range of the category: Category 1 PPE (4 cal/cm² minimum arc rating) provides protection at incident energies up to 4 cal/cm²; Category 2 PPE (8 cal/cm² minimum arc rating) provides protection at incident energies up to 8 cal/cm²; Category 3 PPE (25 cal/cm² minimum arc rating) provides protection at incident energies up to 25 cal/cm². The category thresholds are calibrated to these EBT ratings. An adversarial pixel perturbation that shifts the rendered PPE category classification from Category 3 to Category 2 at a 480 V switchgear position with 18.4 cal/cm² calculated incident energy places a worker in PPE rated for 8 cal/cm² in an arc flash exposure of 18.4 cal/cm² — 2.3× the PPE’s EBT. The failure is immediate and irreversible: the arc flash event, which develops its peak energy deposition in the first 10–50 milliseconds, completes before any human detection, alarm response, or protective action can interrupt the exposure.

IEEE 1584-2018 arc flash calculation methodology and the AI rendering boundary

The IEEE 1584-2018 arc flash calculation model was developed from an empirical database of 1,980 arc flash tests conducted at five laboratories across a voltage range of 208 V to 15 kV, electrode configurations including open-air, LV switchgear, MV switchgear, and MCCs, and a range of gap distances (G) and available bolted fault currents. The 2018 model substantially revised the 2002 edition’s simplified single-equation approach, replacing it with a system of regression equations that account for voltage, electrode gap, equipment type (equipment type factors), and distance correction factors that better capture the geometry-dependent energy distribution in the arc flash plasma. The key inputs to any IEEE 1584-2018 calculation at a specific equipment position are:

System voltage (V) determines which set of regression coefficients applies (the model uses different coefficient sets for the 208–600 V range, the 600 V to 15 kV range). Available bolted fault current (I_bf) — the maximum short-circuit current available at the equipment bus, determined from a system-wide short-circuit study — drives the arcing current calculation. The arcing current (I_arc) is calculated from I_bf using regression equations that also depend on voltage and electrode gap; I_arc is always less than I_bf because the arc impedance reduces the available current. The electrode gap (G, in mm) affects the arc plasma column geometry and thus the spatial energy distribution. Working distance (D, in mm) — standardised as 455 mm for LV panelboards and MCCs, 610 mm for LV switchgear, 910 mm for MV switchgear — is the distance from the arc source to the worker’s face and chest. Arcing time (t, in seconds) is determined by reading the time-current characteristic (TCC) of the upstream overcurrent protective device at the calculated I_arc: if the upstream breaker clears the calculated arcing current in 0.1 s, the arcing duration is 0.1 s; if the breaker takes 0.4 s to operate (as in a long-time-delay overcurrent trip unit characteristic at low-multiple fault currents), the arcing duration is 0.4 s. Arcing time is the variable with the greatest impact on calculated incident energy: doubling the clearing time doubles the incident energy, all else equal.

The arc flash study software — running the full IEEE 1584-2018 model for every equipment position in a facility’s electrical distribution system — produces a comprehensive table of results: one row per equipment bus position, with columns for calculated I_arc, incident energy (cal/cm²), PPE category (per NFPA 70E Table 130.7(C)(15)(a)), arc flash boundary (m), working distance (mm), equipment type, arcing time (s), and study date. This table is typically rendered in the arc flash study software’s report display — a formatted HTML or PDF table with colour-coded category indicators (Category 1 green, Category 2 yellow, Category 3 orange, Category 4 red) and incident energy values displayed in large, readable font for field use. It is also stored as a label data export, populating the physical arc flash hazard warning labels attached to each piece of equipment.

The adversarial injection surface is at the boundary between this rendered table display and the AI classifier that reads it. AI systems from vendors including Schneider Electric (EcoStruxure Power Advisor arc flash module AI), Eaton (Power Xpert arc flash management AI), and SKM Systems (PTW arc flash report classification AI) ingest rendered images of the arc flash study report display to classify the current PPE requirement at a specific equipment position. The AI does not recompute the IEEE 1584-2018 calculation; it reads the rendered output of that computation from the display. A ±8 DN perturbation at the pixel region encoding the category colour and incident energy value — shifting orange (Category 3, 18.4 cal/cm²) to yellow (Category 2, displayed as 6.2 cal/cm²) — causes the AI to classify the Category 3 position as Category 2. The underlying arc flash study database is unmodified; the underlying IEEE 1584-2018 calculation is correct; only the rendered pixel output of the report display has been perturbed. An arc flash study update review, equipment database audit, or physical label inspection would not identify the adversarial manipulation, because the manipulation exists only at the AI’s rendered-image input boundary.

Four adversarial injection surfaces in arc flash incident energy AI

1. PPE category calculation display AI (Schneider Electric EcoStruxure Power Advisor AI, Eaton Power Xpert arc flash AI, SKM Systems PTW arc flash AI, ETAP arc flash AI — rendered arc flash study report table classification AI)

The primary adversarial injection surface in the arc flash AI pipeline is the rendered arc flash study report table — the software display that presents the IEEE 1584-2018 calculation results for each electrical equipment position in the facility distribution system. For a 480 V LV MCC position with 50 kA available fault current and a 0.4 s upstream breaker long-time-delay trip characteristic: I_arc (calculated) → 38.7 kA; incident energy (calculated at 455 mm working distance) → 18.4 cal/cm²; PPE category → Category 3 (above 8 and at or below 25 cal/cm²); arc flash boundary → 3.4 m. The report display renders Category 3 in orange with the incident energy value 18.4 displayed in an orange-bordered table cell. The AI classifier processes this rendered display to approve energised work permit PPE requirements.

A ±8 DN downward perturbation applied to the pixel region encoding the PPE category cell and incident energy value shifts the orange Category 3 cell to a yellow-spectrum encoding indistinguishable from a Category 2 cell in the AI’s training distribution; the incident energy value 18.4 is rendered with perturbed pixel values that the AI reads as 6.2 (a plausible Category 2 value). The AI classifies the equipment position as Category 2, minimum arc rating 8 cal/cm². The energised work permit is approved for the worker’s Category 2 arc flash kit (8 cal/cm² arc flash jacket, 8 cal/cm² arc flash bib overalls, 8 cal/cm² arc flash face shield, 4 cal/cm² arc-rated gloves). The worker approaches the 480 V MCC panel to perform an energised switching task — racking in a new draw-out motor starter while the bus is energised — in PPE rated for Category 2 at a Category 3 position.

When an arcing fault initiates during the racking operation (an arcing fault probability for draw-out motor starters during racking operations of approximately 10–15% per NFPA 70E Informative Annex D), the arc flash event deposits 18.4 cal/cm² at the worker’s face and body at 455 mm working distance. The Category 2 arc flash jacket and bib overalls (EBT 8 cal/cm²) develop breakopen holes in the first 50 milliseconds; the cotton base layers beneath ignite; third-degree contact burns occur across the exposed torso, forearms, and neck. NFPA 70E-2021 Section 130.5(C)(1) permits use of the incident energy analysis method with an AI-assisted energised work permit system — but specifies no adversarial robustness requirement for the AI that classifies the rendered incident energy report display. See the full arc flash incident energy thermal camera AI prompt injection technical specification for the classification boundary parameter detail.

2. Arc flash thermal camera incident energy monitoring AI (Flir T-Series arc flash thermal AI, FLIR A400 continuous monitoring AI, Fluke Ti480 PRO arc flash thermal AI — thermal camera image classification AI at switchgear and MCC positions)

Thermal camera systems deployed at medium-voltage (MV) switchgear (4.16 kV, 13.8 kV, 34.5 kV) and LV motor control centres provide continuous monitoring of bus compartment temperature distributions through IR-transparent infrared windows installed in the switchgear enclosure panels. Normal thermal images show the conductor and contact temperature distribution within the operating temperature rise envelope — typically 20–40°C above ambient — with no localised hot spots above the 70°C copper conductor temperature limit established in NFPA 70B (Recommended Practice for Electrical Equipment Maintenance). A developing arc flash precursor — contaminated bus surface with tracking current, deteriorating cable lug connection with elevated contact resistance, loose bus joint with mechanical creep in service — produces a localised hot spot above 70°C that appears in the thermal image as a red or white region with a temperature overlay value.

AI systems classify these thermal images to determine whether a hot spot is present, characterise its severity against NFPA 70B temperature categories, and assess whether the hot spot represents an elevated arc flash risk requiring PPE category upgrade or equipment de-energisation for inspection. A ±8 DN downward perturbation applied to the pixel region encoding the hot-spot colour in the rendered thermal camera image shifts the apparent hot-spot temperature from the critical red-white zone (130–160°C, “critical” under NFPA 70B severity categories, requiring immediate corrective action) to the yellow-green zone (40–60°C, “minor,” no immediate action required). The thermal camera AI classifies a 13.8 kV switchgear position with a deteriorating bus insulator with surface tracking — generating a 145°C hot spot at the tracking current path — as “minor temperature rise, within normal operating limits, no PPE upgrade required.”

The electrical worker assigned to perform an energised switching operation at that 13.8 kV switchgear bay dons the PPE category specified in the arc flash study label (Category 3 for the calculated incident energy at normal operating conditions) without the PPE upgrade to Category 4 that the hot spot precursor condition warrants. The deteriorating insulator surface tracking condition represents an elevated arc flash probability — above the base-rate probability assumed in the IEEE 1584-2018 incident energy calculation — because the tracking current can initiate an arc flash at a lower-energy electrical event than the full bolted fault calculation assumes. When the arc flash event occurs at this elevated-probability precursor condition, the incident energy may be higher than the Category 3 study value if the arc path evolves through the tracking insulator surface rather than the standard test geometry. Category 3 PPE (25 cal/cm² rated) is sufficient for the nominal calculated condition; for the elevated precursor condition with abnormal arc path geometry, it may not provide the designed safety margin.

3. Flash protection boundary display AI (ABB Ability EDCS arc flash boundary AI, Siemens Spectrum Power arc flash boundary AI, GE Grid Solutions PowerOn arc flash boundary AI — arc flash boundary diagram classification AI)

The arc flash boundary — defined in NFPA 70E-2021 Section 130.5(B) as the distance from the arc source at which incident energy equals 1.2 cal/cm² — establishes the minimum exclusion zone for unprotected workers and bystanders during an energised electrical task. At a 480 V LV switchgear position with 65 kA fault current and a 0.1 s upstream breaker clearing time, the calculated incident energy is 6.1 cal/cm² at 610 mm working distance and the arc flash boundary (where incident energy equals 1.2 cal/cm²) is approximately 1.8 m. At a 13.8 kV MV switchgear with 20 kA fault current and 0.5 s backup clearing time (backup protection operating), the incident energy at 910 mm is 32 cal/cm² and the arc flash boundary is approximately 6.1 m — a substantial exclusion zone in a substation bay environment where other workers, instrument technicians, and operators may be working in adjacent bays.

AI systems classify rendered arc flash boundary diagrams — graphical overlays on facility floor plans or switchgear bay elevation drawings showing the arc flash boundary as a circle or arc around each equipment position — to determine whether bystanders and adjacent workers are within or outside the established exclusion zone before authorising energised work. A ±10 DN perturbation applied to the pixel region encoding the arc flash boundary radius in the rendered floor plan overlay compresses the apparent boundary from 6.1 m (the correct MV switchgear boundary, rendered as a large orange circle) to 2.1 m (a significantly smaller circle, corresponding to a much lower-energy LV equipment boundary). The AI classifies a bystander standing 3.5 m from the MV switchgear bay as “outside the arc flash boundary,” not requiring relocation or arc-rated PPE. The energised work permit is approved. When an arc flash event occurs at the MV switchgear position, the bystander at 3.5 m from a 32 cal/cm² arc source receives approximately 3–4 cal/cm² incident energy — above the 1.2 cal/cm² onset-second-degree-burn threshold — on unprotected skin. NFPA 70E-2021 Section 130.7(A)(1) requires all persons within the arc flash boundary to wear PPE appropriate to the hazard exposure — but the AI classification of the rendered boundary diagram that enforces this requirement has no adversarial robustness specification.

4. Incident energy trend display AI (Schneider Electric EcoStruxure Power monitoring AI, Eaton Power Xpert trend AI, ABB Ability Power Management AI — arc flash study trend classification AI tracking PPE category changes over time)

Arc flash study data is not static: equipment modifications, protective device setting changes, utility system changes, and load growth alter the available fault current and protective device characteristics that drive the IEEE 1584-2018 calculated incident energy at each equipment position. Electrical facilities are required under NFPA 70E Section 130.5(G) to review and update their arc flash risk assessments when major modifications or renovations are made to the electrical distribution system — and industry practice under the NFPA 70E technical committee guidance is to update the arc flash study at least every 5 years or when significant changes occur. Power monitoring AI systems track incident energy trends over time — comparing successive arc flash study results for the same equipment position and flagging positions where calculated incident energy has increased above the current PPE category threshold. A 480 V switchgear position previously calculated at 6.8 cal/cm² (Category 2) that, after a utility upgrade that increased available fault current by 15%, now calculates at 8.7 cal/cm² (Category 3) requires a PPE category update before the next energised work task at that position.

A ±8 DN perturbation on a rendered incident energy trend display that suppresses a rising incident energy trajectory — normalising an upward trend from 6.8 cal/cm² → 7.9 cal/cm² → 8.7 cal/cm² over three successive study updates to a flat line at 6.5–7.0 cal/cm² — causes the trend AI to classify the Category 2-to-Category 3 transition as normal variation within the Category 2 range. No PPE category upgrade alert is generated. Workers continue performing energised tasks at the affected position in Category 2 PPE as the actual calculated incident energy exceeds 8 cal/cm² — above the Category 2 EBT — at every subsequent task. The trend AI adversarial surface is distinct from the immediate PPE category display AI surface: it suppresses the institutional knowledge update that would trigger a PPE category review, rather than misclassifying the PPE requirement at a single task execution. The consequence propagates across all future energised work tasks at the affected position until an unmanipulated arc flash study review corrects the record.

NFPA 70E-2021, IEEE 1584-2018, and OSHA 29 CFR 1910.333: the regulatory specification and the adversarial robustness gap

The United States arc flash electrical safety regulatory framework operates across three interlocking instruments. NFPA 70E-2021 (Standard for Electrical Safety in the Workplace, 2021 edition) is the primary performance standard: Section 130.5 requires arc flash risk assessment for all energised work; Section 130.7 specifies PPE requirements and the Table 130.7(C)(15)(a) category system; Section 130.2(A)(4) establishes the energised work permit requirement for all electrical work on live parts above 50 V at or above 50 V ac/120 V dc. IEEE 1584-2018 provides the calculation methodology: it is not itself a mandatory standard (it is not cited in OSHA regulations) but is the accepted industry method for the incident energy analysis method permitted under NFPA 70E Section 130.5(C). OSHA 29 CFR 1910.333 (Safety Requirements for Working on or Near Exposed Electrical Parts) is the federal regulatory mandate: Section 1910.333(a)(1) requires that live parts to which an employee may be exposed be de-energised before work begins unless de-energising introduces additional hazards or the task is inherently infeasible for de-energised execution; for permitted energised work, 1910.333(b)(2) requires appropriate protective equipment. OSHA enforcement of arc flash fatalities references NFPA 70E as the recognised standard of practice under the General Duty Clause (OSH Act Section 5(a)(1)) when 1910.333 does not specify a precise requirement for the relevant condition.

The adversarial robustness gap is structural across all three instruments. NFPA 70E-2021 Section 130.5 requires that the arc flash risk assessment be performed and documented, and that PPE be selected based on the assessment — but does not specify how the arc flash study report display is protected from manipulation between its generation (by the IEEE 1584-2018 calculation software) and its consumption (by the AI or human classifier determining the PPE requirement at the work task). IEEE 1584-2018 specifies the calculation methodology and input parameters — it is not a standard for the display systems or AI classifiers that process the calculation’s output. OSHA 29 CFR 1910.333 and the General Duty Clause enforcement framework evaluate whether the employer performed an adequate arc flash risk assessment and selected appropriate PPE — they do not address adversarial manipulation of the AI systems that implement those assessments. The adversarial injection attack therefore operates in a regulatory void between the calculation standard (IEEE 1584-2018), the safety standard (NFPA 70E-2021), and the enforcement standard (OSHA 1910.333) — none of which addresses the AI classification layer at the rendered-display boundary.

The structural pattern is consistent with the broader regulatory gap documented across safety-critical AI domains in this portfolio. In oil refinery process control, OSHA PSM 29 CFR 1910.119 requires process hazard analysis and mechanical integrity programs but specifies no adversarial robustness criterion for APC AI classifying rendered process variable displays. In nuclear power plant I&C, NRC 10 CFR Part 50 Appendix A GDC 13 and IEEE Std 603-2018 specify instrumentation and display requirements but not adversarial robustness for AI classifying rendered reactor protection system parameter displays. The arc flash context shares this pattern with a specific additional feature: unlike the nuclear I&C context (where NRC GDC 20–24 require redundant automatic protective action systems independent of the I&C display), the arc flash PPE selection decision has no independent automated interlock. The energised work permit AI is the sole barrier between an adversarially perturbed PPE category classification and the worker approaching energised equipment in underrated PPE.

ESFI arc flash statistics and the consequence envelope for PPE misclassification at scale

The Electrical Safety Foundation International (ESFI) tracks arc flash fatalities and injuries in the United States using Bureau of Labor Statistics (BLS) Census of Fatal Occupational Injuries (CFOI) and Survey of Occupational Injuries and Illnesses (SOII) data. ESFI’s published statistics document approximately 400 arc flash fatalities per year and approximately 2,000 arc flash burn injuries requiring hospital treatment per year across all US industries. The electrical utilities sector (OSHA 29 CFR 1910.269 covered), industrial manufacturing, and construction account for the majority of arc flash fatalities; the electrical utilities sector has the highest incident energy exposures (transmission substation switching at voltages of 115 kV to 765 kV, with ANSI/IEEE C37.20 switchgear and IEEE 1584-2018 calculations producing incident energies of 20–40+ cal/cm² at maintenance working distances).

The ~400 fatalities/year rate means that, statistically, more than one arc flash fatality occurs per day in the United States across all industries. NFPA 70E technical committee materials and OSHA fatality case databases document the distribution of proximate causes: absence of arc flash risk assessment (workers unaware of PPE requirement), outdated arc flash study data (equipment changes not reflected in study), incorrect PPE tier selection (Category 2 PPE donned for a Category 3 or higher position), and PPE not worn despite known requirement (human factor compliance failure). The adversarial injection attack against arc flash PPE category display AI replicates the “incorrect PPE tier selection” failure mode — the third category — with the specific property that it is undetectable by the standard corrective measures for that failure mode (arc flash study update, equipment label verification, safety audit of work permit records). The manipulation leaves all underlying data correct; only the rendered display AI input is perturbed.

At scale across an industrial portfolio deploying arc flash PPE category display AI — a large manufacturing facility with 500 switchgear and MCC positions, each with an AI-classified energised work permit system — the adversarial injection surface represents a potential for systematic PPE category misclassification at multiple positions simultaneously. An adversarial perturbation that applies ±8 DN shifts to the rendered Category 3 cell encoding across all Category 3 positions in the arc flash study report display would cause the AI to classify all Category 3 positions as Category 2 until a manual review against the underlying arc flash database was performed. For facilities with quarterly energised switching maintenance schedules, the window between perturbation and detection could encompass multiple energised work tasks at multiple Category 3 positions — each with the same PPE underrating failure mode.

Glyphward threshold 35 for arc flash incident energy AI

Glyphward’s adversarial detection API operates as a pre-classification gate at each rendered-image ingestion boundary in the arc flash AI pipeline: before the PPE category display AI processes each rendered arc flash study report table, before the thermal camera AI processes each thermal image from switchgear bay monitoring cameras, before the arc flash boundary diagram AI processes each floor plan overlay, and before the incident energy trend AI processes each trend display chart. Each rendered image receives a risk score (0–100) in 8–15 ms. At or above threshold 35, Glyphward gates the AI classification and generates an alert that triggers manual verification of the underlying arc flash study database record — the IEEE 1584-2018 calculation output that is not accessible to pixel-level adversarial perturbation, because it is stored in the arc flash study software’s structured database rather than rendered as a classifiable image.

Threshold 35 for arc flash incident energy AI reflects three consequence factors. First, the PPE EBT failure consequence is immediate and irreversible: the 5–200 ms arc flash event completes before any alarm response, human intervention, or protective action can prevent the energy deposition on underrated PPE. Unlike underground mining ventilation AI adversarial injection — where the methane accumulation timeline provides a window between monitoring suppression and ignition event — there is no intervention window between the arc flash event initiation and the PPE EBT failure. Second, there is no independent automated interlock between the AI PPE category classification and the worker’s approach to energised equipment: the energised work permit system that should prevent underrated PPE approach uses the same AI classification as its input; there is no NFPA 70E or OSHA-mandated independent safety layer that would detect the Category 2/Category 3 mismatch before worker exposure. Third, the ~400 US arc flash fatalities/year base rate means that even a small adversarial success rate applied to the PPE category classification decision at scale produces a statistically meaningful additional fatality count above the base rate.

Threshold 35 is consistent with the Li-ion gigafactory electrode coating AI threshold (35 — thermal runaway consequence, no independent automated interlock between coating inspection AI and cell assembly decision, consumer device fleet scale). It is above threshold 30 for underground mining ventilation AI (post-explosion survivor window), offshore jack-up rig structural stability AI (multi-barrier punch-through detection system), and power substation protection relay AI (cascade blackout consequence but no immediate fatality outcome per-event). It is below threshold 40 for pharmaceutical sterile fill-finish vial inspection AI (patient IV injection mortality with no physiological warning before systemic toxicity) and below threshold 25 for nuclear power plant digital I&C AI (NRC GDC 20–24 redundant automatic protective action systems operating independent of the I&C display classifier).

The false positive cost at threshold 35 — a manual verification check against the underlying arc flash study database for a flagged rendered display — adds approximately 30–90 seconds to a work permit review process. The false negative cost — approving Category 2 PPE for a Category 3 arc flash event, resulting in EBT exceeded, PPE burns through, third-degree burns on the worker — is fatal. The proportionality is not close.

Free tier — 10 scans/day, no card required. Submit a rendered arc flash study report display image or PPE category table from your arc flash management software to the Glyphward scanner to generate a baseline adversarial risk score for your arc flash incident energy AI classification inputs.