Your compressor already found the fault.
When AT-1 compresses a vibration tag, it stores a residual — what the fitted model couldn’t explain. A developing bearing defect leaves periodic impulses in exactly that residual. AT-1’s diagnostic add-on reads them to detect and localize the fault and trend its severity — with no extra sensors, no labelled fault data, and no trained ML model.
- 78%
- faults detected (28/36 real CWRU records)
- 0%
- false alarm on healthy baselines (0/36)
- inner 92% · outer 67%
- fault localized to the right component
- zero
- extra sensors, labels, or trained model
Diagnostics that ride on storage you already buy
It's a free byproduct of storage
When AT-1 losslessly compresses a vibration tag it fits a generating model and stores the residual — the part the model can't explain. That residual already exists. We read its periodic structure; you pay nothing extra in compute to get a diagnostic out of data you were storing anyway.
Detects AND localizes — names the part
Characteristic fault frequencies come from the bearing geometry and shaft speed (ball-pass inner/outer race, ball-spin, cage). When periodic structure emerges at one of them, we map it to the physical component — "developing inner-race defect" — not just "something's wrong."
Earlier than the RMS/temperature alarm
Structure appears in the residual while the machine still runs normally and shows no conventional symptom — before an amplitude or temperature threshold trips. A severity trend across snapshots tells you whether it's getting worse, so maintenance can prioritize.
Validated on real bearing faults
Reduced to practice on 40 real Case Western (CWRU) drive-end recordings — healthy baselines and seeded inner-race, outer-race and ball faults across four loads. Produced label-free, with no training, straight from the compressor’s stored residual.
| Machine state | AT-1 decision | Localization |
|---|---|---|
| Healthy baseline | nominal — no alert | 0/36 false alarms |
| Inner-race defect (0.007–0.021") | ALERT · inner race | localized 11/12 (92%) |
| Outer-race defect (0.007–0.021") | ALERT · outer race | localized 8/12 (67%) |
| Ball defect (0.007–0.021") | weak — reported honestly | localized 2/12 (17%) |
We don’t overclaim: inner- and outer-race faults (the dominant industrial failure modes) are detected and localized strongly; ball faults — universally the hardest mode — are weak, and we report that openly. The method explicitly does not predict remaining-useful-life or time-to-failure, which is not reliably inferable from this signal.
Beyond bearings — any multi-sensor asset
Bearings have known fault frequencies; most assets don’t. For engines, pumps, turbines and whole fleets, AT-1 also produces a condition-agnostichealth index — how far a recent multi-sensor window has drifted from the asset’s own healthy baseline, measured by compression distance. No fault frequencies, no labels, no per-asset tuning.
Validated on NASA C-MAPSS run-to-failure
709 turbofan engines run until they fail. The label-free health index tracks each engine’s remaining life — and no RUL labels were used to build it.
- 0.88
- median rank-correlation vs remaining life (no labels used)
- 709
- real run-to-failure engines, 4 fault modes
- beats naïve
- on multi-condition fleets (0.66 vs 0.24)
- zero
- fault frequencies, labels, or training
The differentiator: it survives real operating conditions
On clean, single-condition data a trivial method ties it — we say so. But on realistic fleets running under changing operating conditions, the compression health index (0.66) crushes residual-magnitude (0.24) and the best single sensor (0.33–0.54): naïve methods collapse when conditions shift; this one doesn’t. That robustness is the edge.
at1 condition-monitor report sensors.csv at1 condition-monitor validate # reproduce the proof
Licensed engine — enable it at /engines, then at1 condition-monitor … works once the at1-engines wheel is installed.
Honest scope: a label-free degradation / remaining-life-rankindicator — a strong “which asset to inspect first” signal — not a calibrated time-to-failure number.
From detect to predict — and triage the whole fleet
The same model the compressor fits to your signal can be run forward to forecast where a channel is heading — and one label-free index can rank every asset you own so a specialist looks at the worst first.
Guarded forecasting
Run the discovered model past the data you have to forecast a trend channel and predict when it will cross a threshold. It is guarded: it abstains — returns an honest “unknown” — rather than extrapolate through a regime change or an oscillation. Validated across 100 NASA C-MAPSS engines: it forecast 97% of units and beat naïve persistence on 89% of them (median skill +0.36).
at1 diagnose forecast sensor.csv --threshold 1400
Honest scope: a trajectory + threshold-crossing forecast on smooth/degradation channels — not a calibrated remaining-useful-life number.
Fleet health screen
Point it at a folder of per-asset signals and it ranks the fleet by a single label-free health index — drift from a known-good baseline, or outlier-versus-peers — flags the worst, and emits CMMS work orders for them. One number, every asset, no per-asset model.
at1 diagnose fleet ./assets --baseline good.csv --cmms work.json
Honest scope: a screen to prioritise specialist attention — not a diagnosis, and not a replacement for specialist analysis.
One probe, many signals — and an honest map of where it works
The same compression-structure probe runs on any signal with no feature engineering. We tested it against the specialist hand-built marker for each domain, on real public data, with confound controls and calibrated nulls. It is strongest on inter-beat cardiac rhythm; we report where it doesn’t win too.
| Signal (public data) | Probe | Specialist marker | Verdict |
|---|---|---|---|
| ECG / atrial fibrillation (RR intervals) | 0.896 | CoV 0.914 | ✓ matches — held-out, zero domain knowledge |
| Sleep apnea (RR intervals) | 0.819 | VLF 0.835 | ✓ matches — confound-controlled |
| Bearing fault (vibration) | NCD 0.770 | kurtosis 0.908 | fires (p<0.001), specialist wins |
| Gait / Parkinson's (stride intervals) | 0.46 | CoV 0.65 | honest null (perm-p 0.67) |
Honest scope: a general model-free screen— strongest on cardiac inter-beat rhythm (it rediscovers the specialist marker with no feature engineering), a useful first pass on vibration, and we say plainly where it’s null. It is not a diagnosis and does not replace a domain-tuned model where one exists.
How it lands in your stack
The diagnostic runs on-prem, on the residual of tags you already store. It emits ready-to-file work orders into your CMMS — the reliability engineer’s tool, a different (bigger) budget line than storage.
Store
You already tier cold vibration/sensor tags to AT-1 for the storage saving. Nothing new to deploy on the sensor side.
Diagnose
Each tag's stored residual is scored for periodic structure at the asset's fault frequencies — label-free, no training corpus.
Alert
A developing defect emits a work order — component, frequency, severity, trend — ready to POST to Maximo, SAP PM, or a webhook.
Prioritize
A rising severity trend across snapshots ranks which machine to inspect first. (Detection + localization + trend — not a time-to-failure estimate.)
We’ll flag faults during your storage pilot
Bring a historian export of rotating-machinery tags. While we prove the storage saving, we’ll also flag any developing bearing faults in that export — on-prem, label-free, free to evaluate.
Patent pending (US provisional filed 2026). Billed as a per-monitored-asset add-on on top of storage.