Skip to content
Ordanis
PT

Guide · Predictive maintenance

Predictive maintenance without a crystal ball.

Predictive maintenance means acting before a failure based on the asset's condition, not only on the calendar. Ordanis starts with the verifiable step: detecting condition deviations in machine signals, with evidence, and measuring how early they showed up.

The approach

  1. 01

    Baseline

    For each signal (temperature, vibration, current…), Ordanis computes the asset's normal running behaviour and records the baseline's confidence.

  2. 02

    Detection

    Configurable rules compare the data with limits and with the baseline. Only good-quality data counts.

  3. 03

    Finding with evidence

    Every finding carries the data window, the rule that fired, the confidence and the signals involved.

  4. 04

    Human decision

    A person reviews the recommendation, approves or rejects it, and the action becomes a tracked work order.

  5. 05

    Measurement

    When the failure happens (or is avoided), Ordanis measures the detection lead time and records the outcome.

Detection rules available

  • 01

    Threshold with hysteresis

    Fires when a limit is crossed and only re-arms below the return band — no alert bursts.

  • 02

    Baseline deviation

    The window mean must deviate beyond a percentage or a score for several consecutive windows, only while the asset runs.

  • 03

    Rate of change

    Detects rises or drops too fast for the signal's normal behaviour.

  • 04

    Flatline

    A value that does not change for too long usually means a sensor or communication problem.

  • 05

    Missing data

    Missing data is a finding in itself: without data nobody can claim everything is fine.

  • 06

    Composite rule

    Combines rules within a time window, e.g. temperature and vibration above baseline within 30 minutes.

How we measure whether it worked

Detection lead time
Version: lead-time
t(reference event) − t(first related detection)

How long before the failure (or reference event) the first related detection appeared. It is the honest measure of a predictive pilot.

What this is not

  • It is not a remaining-useful-life (RUL) estimate or a failure-date forecast.
  • It is not a machine-learning model trained on your plant's failures.
  • It promises no downtime reduction: pilot targets are only set after the baseline with your data.
  • It never writes to a PLC or acts on its own: the machine control policy is read-only.
  • It has not been validated with a customer's production data yet; tests use a simulator.
Ordanis

One Industrial Reality.

Start with one operational problem. Expand into one industrial reality.

Request Pilot