Docs · patent pending

Condition monitoring: method & results

AT-1’s diagnostic add-on reads the residual a lossless compressor already produces and turns it into a bearing-fault diagnostic — detecting and localizing a developing defect and trending its severity, with no separately trained model, no labels, and no training corpus. This page documents the mechanism, the verified results, the alert format a CMMS consumes, how it’s billed, and the boundary where the method stops.

What the diagnostic does

To compress a vibration time-series losslessly, AT-1 fits a generating model (in the preferred embodiment an autoregressive / linear-predictive model) and stores the model parameters plus an entropy-coded residual— the exact difference between the data and the model’s quantized prediction. For a healthy machine that residual is essentially structureless. A developing defect — a spall on a race, a pit on a ball — produces a periodic impulse every time a rolling element strikes it, and those impulses survive into the residual.

AT-1 computes an envelope spectrumof the residual (a Hilbert envelope, then an FFT) and scores the magnitude at each of the bearing’s characteristic fault frequencies — relative to a local median. A healthy residual scores near baseline; a developing fault scores tens-to-hundreds at its own frequency. Because each fault frequency maps to a physical component, the fault is localized, not merely detected.

Characteristic fault frequencies

Computed from bearing geometry and shaft speed. For a rolling-element bearing these are the ball-pass inner-race (BPFI), ball-pass outer-race (BPFO), ball-spin (BSF), and cage (FTF) frequencies. The defaults below are the CWRU SKF 6205-2RS multipliers (× shaft speed in Hz); override per your bearing.

BPFI = 5.4152 · fr     # inner race  -> "bearing inner race"
BPFO = 3.5848 · fr     # outer race  -> "bearing outer race"
BSF  = 2.3570 · fr     # ball / roller -> "rolling element"
FTF  = 0.3983 · fr     # cage        -> "cage / train"
   ( fr = rpm / 60 )

Verified results (real CWRU bearing data)

Reduced to practice on 40 real Case Western Reserve drive-end recordings — healthy baselines plus seeded inner-race, ball and outer-race faults at 0.007 / 0.014 / 0.021″ across four loads. Each recording was compressed and its residual scored, label-free.

Inner- and outer-race faults — the dominant industrial failure modes — are detected and localized strongly. Ball faults are universally the hardest mode and the method is weak there; we report it openly rather than hiding it.

The alert / work order

A flagged asset emits a CMMS-ready work order — field names mirror the common Maximo WORKORDER / SAP PM notification shape, so it drops onto a connector with no mapping:

{
  "asset_id": "PUMP-101.DE",
  "location": "Plant A / Line 2",
  "problem_code": "BPFO",
  "description": "Developing bearing outer race defect detected at 107.4 Hz
                  (structure score 31.2, elevated, worsening). Label-free
                  residual diagnostic; inspect before the RMS/temp alarm.",
  "priority": 2,
  "severity": "elevated",
  "trend": "worsening",
  "detected_freq_hz": 107.4,
  "source": "AT-1 diagnostic-residual",
  "note": "Detection + localization + severity trend. NOT a remaining-useful-life estimate."
}

A severity trend (worsening / stable / first-seen) is computed across successive snapshots so maintenance can prioritize the worst machine first.

Billing

The diagnostic rides the same metering rail as the AT-1 encoder. Each monitored asset is registered as an inventory line (archive_id="diag:<asset>") and the analyzed signal volume is reported on the I/O axis. It’s sold as a per-monitored-asset add-on on top of a storage deal — the reliability-engineering budget, not the storage admin’s. Offline / unlicensed use is a no-op and never blocks.

Condition monitoring is a licensed engine. Enable it from your dashboard at /engines (requires a card on file), accept the EULA, and download the wheel — then at1 condition-monitor report sensors.csv works once the at1-engines wheel is installed.

The honest boundary

This is detection + localization + a severity trend, earlier than a conventional RMS-amplitude or temperature alarm while the machine still runs. It is not a remaining-useful-life or time-to-failure prediction — that is not reliably inferable from this signal, and we tested for it explicitly (a multi-bearing run-to-failure study gave only 2/4 robust early warnings, so we dropped the RUL claim). The value is catching the developing defect, and naming the part, with zero extra sensors, ML, or labels.

See the condition-monitoring overview, the underlying generative compression tier the residual comes from, and request a pilot.