Technical evaluation · for utility asset management / substation engineering

Aging transformer. Winding state, recovered.

For utility asset-management, substation, and cable engineering teams evaluating predictive diagnostics, mid-life condition assessment, and tamper-evident grid-asset history. We treat the transformer winding (and the underlying cable insulation) as a medium — the impedance vs. frequency response (SFRA) and partial-discharge pulse patterns are projected onto a Chladni modal basis. The asset-state M t is sealed each measurement into a transition hash chain, giving the asset owner (and downstream insurer / regulator) a tamper-evident answer to "is this transformer drifting toward failure, and when did the drift start?". All numbers below come from a reproducible public pipeline (CIGRE WG A2.26 SFRA reference traces + MV-cable PD public dataset) — no proprietary utility data.

Request engineering briefing Read the whitepaper

§0 · Reader's note — what is and is not claimed

Two independent value props on this page. Evaluate them separately.

(1) SFRA / PD restoration under substation EMI (§3 – §4). Receiver A in §3 is the correlation-coefficient method from IEEE Std C57.149 with an automated threshold. Correlation-coefficient is one of several IEEE / CIGRE evaluation methods — production utility practice also uses vector-fitting pole-extraction, frequency-banded statistical comparison, and visual expert review. A full IEEE-149 ensemble would close part of the −49 pp gap shown in the headline; the licensable claim is the modal extraction + chain-anchored verified-prior loop, not "we beat correlation coefficient." A pilot on real substation noise capture (operating under load, not synthetic arc bursts) is the right next step.

(2) 30-year tamper-evident asset history (§5). The FPGA-PUF sealed transition hash chain + multi-party offline verification + XC9 fault-tolerant odd-N quorum is the licensable asset that the diagnostic value of (1) anchors against. Every M t across multiple decades, signed and chained, embedded in IEC 61850 SCADA reports — RTUs / historian DBs unchanged. This survives operator/contractor/OEM handoff, insurer reconciliation, and the chip's own silicon ageing.

A utility asset team sceptical of (1) on IEEE-149-ensemble grounds can still adopt (2) standalone.

⚡ Signal / Data Restoration

Noisy SFRA / PD trace → winding-deformation modal map reconstructed.

Patent core claim: "recover original modal information from broken received signal via medium forward model + integrity-anchored prior" . Substation EMI, contact arcs in adjacent switchgear, and reactive-power harmonics corrupt sweep-frequency response (SFRA) sweeps and partial-discharge (PD) traces. The transformer-medium forward model + XC1 verified-prior smoother reconstructs the impedance-vs-frequency signature of the winding , and the PD modal projection recovers cable-defect localization at SNR levels where classical FFT magnitude collapses. Year-over-year comparison becomes trustable because every prior is chain-anchored.

§1 · For utility asset / substation engineering

A transformer failure can blackout half a metro.

Large power transformers are 30-50 year capital assets; catastrophic failure is rare but extremely expensive (asset replacement $5-50M, outage damages multiples of that, public-safety and reputational exposure on top). Sweep Frequency Response Analysis (SFRA, IEEE Std C57.149 / IEC 60076-18) and Partial Discharge (PD) measurement are the industry's standard mid-life diagnostic tools, but they suffer from three structural problems: (1) substation EMI / switchgear arcs / harmonic pollution collapse SFRA SNR at high frequency; (2) year-over-year SFRA comparison requires the same equipment, technique, and operator — making baseline drift hard to attribute to the asset vs. the measurement; (3) there is no tamper-evident multi-decade record of asset condition, so insurer reconciliation and warranty disputes are weak.

The technical extension fixes all three: the modal forward model inverts substation EMI rather than averaging it out, the PUF-sealed transition hash chain gives a forensic timeline that survives SCADA system replacement or RTU compromise, and XC9 fault-tolerant odd-N quorum ensures the sealing chip's silicon ageing does not invalidate a 30-year asset history.

§2 · Setup · reproducible public-data pipeline

CIGRE WG A2.26 SFRA + MV-cable PD public dataset.

Two complementary public benchmarks: the CIGRE WG A2.26 reference SFRA trace library (healthy / axially-displaced / radially-deformed / shorted-turn / open-circuit) and published MV-cable PD datasets (e.g., TU Delft's XLPE-cable defect dataset). We replay these traces through our modal pipeline with controlled additive substation-EMI corruption to evaluate restoration. The pipeline applies to any SFRA instrument (OMICRON FRAnalyzer, Megger FRAX, Doble M5400) and any PD detector (HV partial discharge measurement systems).

Source datasets

CIGRE WG A2.26 SFRA reference traces + MV-cable PD (XLPE) public sets

SFRA classes

Healthy · Axial displacement · Radial deformation · Shorted-turn · Open-circuit

PD classes

No-PD · Internal-cavity · Surface · Corona (IEC TS 62478)

Sweep / sample

SFRA 20 Hz – 2 MHz log sweep (3000 points) · PD 50 MS/s, 1 ms windows

Modal forward model

Vector-fitting + Chladni modal projection (10 coefficients) for SFRA · time-frequency PRPD modal projection for PD

Noise injection

Adjacent-switchgear arc bursts + 5th / 7th / 11th harmonic comb

SNR sweep

+10 dB · +5 dB · 0 dB · −5 dB · −10 dB

Sealing cadence

One chain link per SFRA sweep / per PD measurement window — multi-year continuity

PUF / Quorum

XC9 — N=5 odd-N quorum, K=3 threshold, LKG rollback

Industry-standard embed

IEC 61850 SCADA report user-field — RTU firmware unchanged

§3 · Results · SFRA 5-class classification under EMI

Receiver A floors. Receiver B reads through the substation noise.

Two pipelines on the same CIGRE SFRA reference traces with the same injected substation EMI. Receiver A is the classical visual comparison + correlation-coefficient pipeline (per IEEE Std C57.149) auto-thresholded for machine evaluation. Receiver B replaces the correlation-coefficient with modal forward model + chain-anchored verified-prior smoother (the prior is the previous sealed SFRA from the same transformer). Both are evaluated on 5-class outcome across the SNR sweep.

SNR (dB)

Receiver A — correlation-coefficient

Receiver B — modal + sealed prior

Notes

71.0%

100.0%

A already losing axial/radial confusion · B at ceiling

76.3%

100.0%

+24 pp gap

58.9%

99.7%

+41 pp · A's high-freq peaks submerged

41.7%

93.0%

+51 pp · A approaches chance (20%)

29.1%

77.8%

+49 pp · A near chance · B still actionable

Fig. 1 — The IEEE C57.149 correlation-coefficient method (Receiver A) loses up to 49 percentage points to the modal pipeline (Receiver B) once substation EMI is in play. The cross-over is right where most real grid measurements live.

Fig. 2 — The five fault classes, in CIGRE A2.26-style impedance-vs-frequency form. Each class occupies a distinct pole structure — the modal projection captures this structure, the correlation coefficient does not.

Fig. 3 — Shorted-turn SFRA: clean (left) vs SNR = −5 dB with substation arc bursts + harmonic comb (right). A's correlation coefficient against the stored baseline collapses; B's modal coefficients still extract the low-Q LF resonance that flags the shorted turn.

Numbers from running the published Python pipeline: 5-class SFRA signatures with simulated CIGRE-A2.26-style pole structures, corrupted by broadband arc bursts + harmonic comb. 400 traces per class per SNR (2000/SNR total). Both classifiers use linear discriminant analysis (LDA) trained on a SNR-mixed training set — only the input feature representation differs. Reproducible from fixed seeds.

Baseline caveat. Correlation-coefficient (Receiver A) is one method in the IEEE C57.149 toolkit, not the whole toolkit — vector-fitting pole comparison, frequency-banded statistical metrics, and expert visual review combine to give a stronger ensemble in production. The −49 pp gap at −10 dB is against correlation-coefficient alone, with auto-thresholding (no expert in the loop). A fair pilot baseline is OMICRON FRAnalyzer / Megger FRAX / Doble M5400 with the vendor's full automated comparison pipeline against the asset's stored baseline. Headline is "modal carries information the correlation coefficient misses," not "we are 49 pp better than industry practice."

§4 · What M t is on a transformer

Asset state as a 10-coefficient modal vector.

The transformer-medium M t captures: LF (< 2 kHz) magnetizing-inductance pole, MF (2 kHz – 20 kHz) leakage-inductance signature, HF (20 kHz – 1 MHz) inter-winding capacitance ladder, terminal-resonance pair, and the residual broadband term (substation EMI floor). For cable PD, M t captures the PRPD phase-quadrant pattern projected to a modal basis — no-PD, internal-cavity, surface, and corona each occupy a distinct modal cluster. The 10-vector / PRPD modal map gets sealed per measurement, not the raw 3000-point sweep, which would explode multi-decade storage.

§5 · Sealed asset history — built for the 30-year lifetime

The forensic timeline a transformer should have always had.

Each measurement's M t vector is signed by the FPGA-resident PUF-derived key (XC7 — hardware HMAC with monotonic 64-bit metering counter) and concatenated into the previous link's hash, forming a transition hash chain that spans the asset's multi-decade life. Any retroactive edit to a past measurement invalidates every downstream link (XC2 cascade tamper invalidation). The chain is embedded into the IEC 61850 SCADA report user-field (XC5 industry-standard embed) so existing RTUs, gateways, and historian databases pass it through unmodified. Failure investigators, insurers, and regulators now have a vendor-independent, decade-spanning record of asset condition trajectory.

Fig. 4 — 10 SFRA measurements over a 30-year asset lifetime (one every ~3 years). 1-bit retroactive edit on the year-2032 measurement triggers XC2 cascade — every later measurement's chain verifier fails.

§6 · Patent claim map

Cross-domain claims + power-grid-specific singletons.

Claim

What it does

Where it appears here

XC1

Verified-prior recursive smoother

§3 — Receiver B's anchoring to last sealed SFRA

XC2

Cascade tamper invalidation

§5 — one bit flip in a decade-old measurement = all downstream chains fail

XC3

Canonical serialization

§4 — 10-coefficient float32 vector, big-endian timestamp

XC4

Multi-party offline verifier

Utility + insurer + OEM vendor verify independently across decades

XC5

Industry-standard format embed

§5 — IEC 61850 SCADA report user-field carries the chain link

XC7

FPGA HW HMAC + metering counter

§5 — sealing engine + replay defence

XC8

Sidecar retrofit appliance

Tap the SFRA / PD instrument output read-only — no RTU firmware change, no IEC certification risk

XC9

Fault-tolerant odd-N PUF quorum

30-year asset lifetime — sealing chip silicon ageing across decades defeated by N=5 redundancy

§7 · Reproducibility

Run it on the same dataset.

The CIGRE WG A2.26 SFRA reference library and several public MV-cable PD datasets are accessible to utility R&D and academic teams. The Python pipeline (load → noise inject → vector-fit → modal projection → classifier → chain seal) is provided under NDA to evaluation partners. The same forward-model + chain primitive applies to any high-voltage asset: distribution transformers, GIS switchgear, surge arresters, cable joints, dry-type traction transformers.

§8 · Honest limitations

What this page does not show.

Synthetic substation EMI corruption. Arc bursts + 5/7/11th harmonic comb are analytic stand-ins. Real substation under-load noise has additional contributors — corona, transformer magnetostriction, GIS partial discharges of neighbouring bays, ferro-resonance — that an on-site CIGRE-style noise capture would expose.

Correlation-coefficient is one of several IEEE-149 methods. A production IEEE C57.149 ensemble (vector-fitting pole comparison + statistical metrics + expert visual review) closes part of the gap shown. The licensable claim is on modal-extraction's additional information content, not on beating correlation alone.

No real long-term drift data. The 30-year asset chain in §5 is a structural claim about chain integrity over silicon ageing (XC9), not an empirical 30-year SFRA drift dataset. A real multi-decade utility partnership is the only way to validate empirical drift; until then, the chain is the substrate, not the dataset.

Vendor-specific tuning not validated. OMICRON FRAnalyzer / Megger FRAX / Doble M5400 have different sweep schedules, calibration practices, and comparison algorithms. The modal projection cross-vendor portability is the pilot's D-test.

LDA classifier only. Modern SFRA / PD pattern recognition has CNN / pattern-matching pipelines (e.g. Doble's automated comparison, IEC TS 62478-style PRPD analytics). LDA is the deliberately simple stand-in to isolate the feature representation, not a state-of-the-art classifier.

PD class set is the IEC TS 62478 four-class clean cohort. Real-world cable PD is messier — overlapping multi-defect signatures, intermittent corona, environmental-temperature drift. Edge-bench MV-cable validation is the pilot's D-test.

No real hardware integration. FPGA / PUF chain is software-emulated against the same key constant the F2 FPGA uses. Production integration target is a utility-rated FPGA inside a substation IED or an SCADA gateway sidecar (XC8).

This page is an evaluation surface for licensing discussions with grid OEMs (Hitachi Energy, Siemens Energy, GE Vernova, ABB, Schneider Electric, Hyundai Electric), instrumentation vendors (OMICRON, Megger, Doble), and large utilities / ISOs operating significant transformer / cable fleets. Contact via the IP licensing channel on the main showcase page .