Untrained predictive maintenance — learned from normal, measured on real failure data

We flag machine degradation without any training or labelled fault data. The system learns each machine's healthy “normal” from its own first clean period, then scores how far new data departs from it — across vibration, sound, current, thermal and multi-sensor signals. Below are measured results on open run-to-failure datasets, the sensors it works with, and honest limits. Results shared openly; the mechanism, formula and recipe are never disclosed. Numbers are provisional. Authored by Tetracta AI Teams.

What we do differently

Headline result — bearing prognosis

A real run-to-failure bearing (FEMTO C1, 234 min of life). Our health index (teal) declines smoothly and monotonically toward failure; a classic scalar index (grey) is noisy. The statistical alarm — healthy mean + 5σ with hysteresis, no manual tuning — fires 92% of life before failure.

The raw vibration the system sees — healthy (start) vs end-of-life, same scale:

Across 32 run-to-failure bearings, 2 independent rigs (FEMTO + XJTU), 6 conditions: median degradation monotonicity +0.96 / +0.98 — beating classic RMS/kurtosis indices with Wilcoxon p<0.001. Per-run monotonicity (honest — strong cases and harder ones shown):

Live pipeline — streaming, on a small server

Replaying a real run-to-failure test stream through the production pipeline (NumPy only, no GPU; the simulator sends the same packets a field sensor would): degradation monotonicity 0.994, alarm at ~31% of life (≈69% early-warning lead), ~0.7 ms per snapshot. The same core runs server-side, so the method stays protected while the user sees live health, remaining-life and alarms. See the public demo →

Results at a glance

Every row is a real run on an open dataset — honest numbers, including where we are weak.

Detection & diagnosis — multi-modality

Fault detection strength (AUROC, 1.0 = perfect) from the same untrained core, by modality:

Sound adds a second view (e.g. pump cavitation), thermal is contact-free, and sensor fusion lifts weak single-modality cases. Aircraft-engine remaining-life (NASA C-MAPSS, 21-sensor): degradation monotonicity up to +0.90. Fault-type separation via order analysis (RPM-normalised) is reported honestly as a guide, not a verdict.

By modality & fault — including pump-sound and thermal

For pumps, sound (cavitation/wear) is the decisive modality — our untrained core reaches AUROC 0.77 there, edging a trained autoencoder (0.73). Thermal is contact-free and very strong on misalignment and rotor faults. Fusing modalities lifts weak single-sensor cases (e.g. imbalance 0.95 → 0.93 fused with sound).

Which sensors it works with

We don't sell the sensor — you connect what you have, and the system tells you honestly what it can do with it. Capability depends on the data tier and bandwidth:

Open-IP raw/spectrum sensors and gateways (e.g. Erbessd, Wilcoxon, NCD.io, Treon, IFM, Banner-class) integrate fully; closed-cloud platforms that don't expose their data are not compatible by design.

Honest limits

We'd rather under-claim. The results above are on laboratory / open datasets; a field pilot is the next milestone and the deciding test. Full value needs FULL-tier sensors (raw/spectrum); on scalar-only sensors you get trend, not early diagnosis. The primitive is solid engineering — not a one-of-a-kind formula; the value is the validated, sensor-agnostic, zero-config, honestly-characterised package, deployed and supported. On the most critical assets, a full vibration analyst still has a place; we play the under-monitored long tail and the entry tier.

Methodology & reproducibility

Results are on open, citable datasets — FEMTO PRONOSTIA and XJTU-SY (bearings), MAFAULDA (imbalance/misalignment), NASA C-MAPSS (aircraft engine), public thermal sets. The core is deterministic with zero learned parameters, so a run reproduces bit-for-bit. We guard against fooling ourselves: temporal-placebo controls, leak-free point-in-time evaluation, and we report medians across many bearings rather than a cherry-picked best. Where a result is weak — a hard bearing, or fault-type separation by clustering — we report it plainly. This is engineering, audited like engineering; effects and methodology are shared, the mechanism is not.

For sensor makers & reliability teams

If you build or deploy condition-monitoring sensors and want a zero-training, sensor-agnostic analytics layer behind them — or a white-label / OEM core — we have a public demo (streaming real test data), an SDK/integration path, and a reproducibility evidence chain. We are also looking for field-pilot partners to test on real, in-service machine data. Methodology and results are shared openly; the core mechanism is licensed, not disclosed.

Write to [email protected] →   See the public demo   ← Research Notes

To set up a call or get the technical brief / evidence chain, email [email protected].