An HV transformer FAT is alive or it is nothing — a summarised PDF does not answer the question that arrives 5 years later.

The problem

A FAT that delivers a PDF delivers a commitment, not a dossier.

The FAT of an HV transformer is one of the most complete and expensive tests in heavy industry: ratio, resistance, dielectrics, PD, SFRA, DGA — all of it per tap, per winding, per phase. The classic consequence of “everything checks out” is a summarised PDF of several dozen pages that the customer signs and files. And while everything goes well, it works.

The four questions that break the PDF:

1. “What exactly was this transformer like when it shipped?” — asked by O&M when a DGA alert appears years later. If the raw curve stayed in the lab in an unsealed folder, reconstructing it is archaeological work.
2. “Why does this transformer from the batch give a different SFRA response?” — comparing 10 units from the same batch to find the outlier requires living raw data, not PDFs.
3. “What is the difference between FAT and SAT?” — an honest comparison demands both curves, not two consolidated figures. Today it is done by hand.
4. “Was any deviation approved during the FAT?” — if the signature lived on a sheet of paper, it gets lost.

The sector that knows this pain better than anyone is aerospace. The good news: the solution proven in aerospace fits power transformers.

How it fits the IRIS system

iLEAN Tracer is the filler between the test benches, the lab and the customer's O&M.

The FAT data exists — it is raw inside the test bench, in the insulation analyser, in the DGA system, in the lab engineer's notebook. What is missing is capturing it at second zero and tying it to the serial number of the specific unit, instead of summarising it in a PDF. iLEAN Tracer sits on top of your FAT lab without asking you to change benches, MES or instrumentation suppliers.

Connect reads benches, the engineer's voice and approvals. Edge reads the winding and unit serials. The agent assembles the living FAT per transformer. A person signs — never the other way round.

The iLEAN pieces applied to piece-by-piece FAT:

• Connect on the test benches — captures raw data from dielectric, resistance and SFRA benches (serial port, USB, OPC UA, local file) at second zero, linked to the unit, the test, the tap, the winding and the phase. If the bench is old and only leaves a file on the lab's industrial PC, Connect picks it up and labels it. Deviations approved verbally are recorded with a timestamp and a signature.
• Edge in the lab — a vision terminal that reads the transformer label and the winding labels, making sure the test runs against the correct unit — not against “the next one in the batch”. If the network drops, it keeps reading and recording locally.
• Living FAT agent — builds the per-unit dossier with queryable raw curves: insulation per tap, PD per phase, SFRA per winding, initial DGA. When the question comes from the customer or from post-commissioning O&M, the answer is immediate and comparable. If a pattern appears (an insulating paper batch with higher PD than expected, an oil supplier with atypical initial DGA), the agent flags it. A person signs every conclusion that affects delivery or a field campaign.

See the full IRIS architecture →

Before and after

Summarised PDF FAT vs. living FAT with iLEAN Tracer

Aspect	Classic PDF FAT	With iLEAN Tracer
Test curves	Consolidated summary at the end, raw data left in the bench	Streamed at second zero, alive and queryable
Link to serial number	Manual, risk of mixing units from the batch	Edge verifies unit and winding on every test
Approved deviations	Signature sheet filed on paper	Approval with timestamp and signature in the system
FAT vs. SAT comparison	By hand, weeks, partial	Direct, graphical comparison in minutes
Outlier detection within the batch	Impossible without manual reconstruction	The agent flags responses anomalous for the family
Behaviour if the network drops	n/a	Connect/Edge keep capturing locally

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the actual data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• HV transformer plant with a FAT lab (dielectric, resistance, SFRA, DGA). Connect on the benches + Edge on serial reading as a pilot, tied to one priority transformer family.
• First value expected within a few weeks: a living FAT running for every unit that goes through the pilot bench, demonstrable to the customer and to the EPC.
• Indicative payback between 4 and 9 months, depending on the historical average cost of post-delivery claims whose root cause requires reconstruction and on the weight of premium utility/EPC customers in your portfolio.
• Hard levers: reduction of hours spent reconstructing raw FAT data ≥ 30%; O&M alerts closed with a direct FAT vs. SAT comparison; commercial premiums tied to demonstrable traceability.

And the fair doubt of the FAT lab manager

“What if the AI files a curve against the wrong unit?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI simply recontextualises a piece of data (linking a test to a unit according to the serial read by Edge), the best models brought error below 1.5% ^[1]. And even so, nothing critical is decided alone: the agent proposes, a person signs the FAT. The three security rings exist precisely for this — the lab's OT network is not touched, it is only handed a signed envelope that a human decides to open.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.