What Sentinel actually does in production
// LANE-3 · physics-informedFATHOM Sentinel ingests six classes of signal per monitored asset: dissolved-gas analysis (DGA) where the asset is liquid-immersed, partial-discharge waveforms, top-oil and winding-hot-spot temperature, load history, weather, and crew-reported audible/visual cues from the field maintenance interface. Each signal is weak on its own; together they are decisive. The lane does not pattern-match — it asks "does the physics of this asset class predict this telemetry, given the input window?" — and complains when the answer is no.
For each asset class we maintain a physics-informed autoencoder. The encoder compresses the recent telemetry window into a low-dimensional latent state; the decoder reconstructs it. The reconstruction loss is augmented by a physics residual term: for a transformer, the difference between observed top-oil temperature and the IEEE C57.91 thermal-model prediction given the load and ambient history; for a Li-ion module, the difference between observed terminal voltage and a coupled electrochemical + thermal cell model; for an inverter, the difference between observed AC current harmonics and the modulation-strategy reference; for switchgear, the difference between observed contact-resistance drift and the operating-cycle history.
This is not a marketing claim. It is the architectural reason FATHOM Sentinel can flag a developing turn-to-turn winding fault four to seven weeks before the DGA threshold would trip, and the reason it can flag a cell-level imbalance in a 4 MWh battery rack two weeks before the BMS would isolate it. The lane is not pattern-matching against past failures alone; it is being told what the physics looks like, and complaining when reality deviates.
The model and what it is trained on
// 12,400 assets · 2,180 confirmed failuresThe core model for transformer monitoring is a graph transformer over a small grid neighbourhood (the monitored asset plus its electrical neighbours within two electrical hops), fused with a 1-D convolutional tower for partial-discharge waveform analysis. It was trained on a multi-DNO fault library of 12,400 historical assets including 2,180 confirmed failures, with strict temporal hold-outs by both calendar year and DNO operator to prevent leakage. The same architecture, with class-specific physics residuals and asset-specific encoders, generalises across the five asset classes we currently support.
Crucially the model is trained against the physics residual, not just the reconstruction loss. That single architectural choice prevents the model from learning to "explain away" sensor drift by adjusting its latent state — a failure mode we observed early in development and which renders pure-autoencoder approaches dangerous in production. A Sentinel event is only emitted when the reconstruction is good and the physics residual is bad, which is the only combination that reliably distinguishes a developing fault from a sensor problem.
Calibration was done in two passes: lab calibration against a fault-injection rig built with the partner DNOs (controlled DGA spikes, controlled PD events, controlled thermal excursions), and then a 14-month live trial across two UK DNOs covering 1,842 transformers and 412 battery racks. The 19-day median lead time, the 95% confidence interval of 7–34 days, and the 0.31 false-positives-per-asset-per-year figure are the calibrated production numbers from that trial — not lab numbers.
SENT · MODEL
- ARCH
- graph TX + 1-D conv · PI loss
- TRAIN SET
- 12,400 assets · 2,180 failures
- HOLD-OUT
- by year · by DNO
- WINDOW
- 30 s rolling · 1 hr context
- LEAD TIME
- median 19 d · 95% CI 7–34 d
- FALSE-POS
- 0.31 / asset / yr (target <0.4)
- RETRAIN
- quarterly · with audit diff
What every event carries
// the evidence packThe lane outputs are designed for the maintenance and asset-management workflow, not just the dashboard. Every event carries: a risk score on a calibrated 0–100 scale; a confidence interval from the calibrated quantile of the physics residual; a suggested intervention window driven by the predicted progression curve; the full input telemetry window (30 s rolling, 1-hour context); the reconstruction the model produced; the physics-residual breakdown by component (thermal, electrical, mechanical); the model version that emitted the verdict; and a signed receipt. If your reliability engineer wants to argue with the model, they can — the evidence is in front of them.
Events flow into the customer's existing CMMS over a standard adapter (we support Maximo, Infor EAM, SAP PM and a handful of bespoke maintenance stacks). They also surface as one of the three event triggers for the LANE-2 dispatch recourse — a developing asset fault on a participating battery can pre-empt a wholesale commitment that would have stressed the asset further. The integration is automatic and the operator can disable it asset-by-asset.
The data discipline that makes it work
// what most asset systems quietly deleteSentinel does not work on telemetry that has been thrown away. Stop discarding DGA samples after one read. Stop storing PD waveforms only when a fault is already suspected. Keep top-oil samples at full cadence, not just the daily average. Digitise crew-reported cues — the engineer who wrote "slight humming on TX-308" in a paper log was right two weeks before the BMS knew. The data that makes Sentinel work is data that most asset-management systems quietly delete because nobody has yet shown an operational use for it.
The first thing we do on a Sentinel engagement is a data-discipline audit: what is being collected, what is being retained, what is being thrown away, what the storage cost would be of keeping it, and what the predictive value of keeping it is. The output is a one-page recommendation; we have published the structure as a checklist in the docs section so that operators who are not yet ready to engage can run the same audit themselves.
Continue exploring
// the other two lanesFATHOM Forecast — LANE-1. The probabilistic load & generation forecaster. Its calibrated quantile bands flow into LANE-2 and, increasingly, into Sentinel's contextual feature set (a forecast of asset stress drives the model's expected reconstruction).
FATHOM Dispatch — LANE-2. The co-optimisation lane. Sentinel events are one of the three event triggers for Dispatch's five-minute recourse cycle, allowing a developing asset fault to pre-empt revenue-stack decisions that would otherwise have stressed the asset further.
Ready to evaluate FATHOM Sentinel?
The standard pilot is twenty assets of a single class — typically transformers on a single primary substation — four weeks of shadow-mode reading your existing telemetry, then a side-by-side report against your current condition-monitoring tooling. We do not charge for shadow-mode.