Skip to content

Condition-Based Monitoring

Turn machine readings into health scores, alerts, trend forecasts and remaining-useful-life estimates, and check the predictions against real failures.

Updated

Watch it in the product

See how OpexMX reads vibration spectra, classifies each machine against ISO 20816 zones, checks bearing fault frequencies and opens a work order with the evidence attached.

Find it: Condition Monitoring in the sidebar (Insights section): Overview, Alerts inbox, Outcomes, Models & Evaluation, Vibration Monitoring, Trend Prediction, Pilot Selection, Data Quality.

Predictive maintenance overview dashboard with asset health scores, anomaly detection status, and remaining-useful-life indicators.

Sound familiar?

  • "Why did that motor burn out? There was no warning."
  • "We replace bearings on a fixed schedule. Some are still good, some fail early."
  • "The PLC has all this data. Can't we use it?"

Condition Monitoring takes the readings your machines already produce (temperature, vibration, pressure, current, speed) and turns them into a health score per asset, alerts that land in one inbox, forecasts of when a value will cross its limit, and an estimate of remaining useful life. Every alert can become a ticket, and every completed repair can be labelled, so you can measure how well the predictions matched what actually failed.


Getting Data In

Readings arrive as asset parameter values. The sources are:

SourceTypical use
OpexFlow edge connectorPLCs, SCADA, CNCs and IoT sensors over Modbus, OPC-UA, EtherNet/IP, MQTT, MTConnect or Fanuc FOCAS, sent in batches
Vibration edgeA vibration sensor on a rotating machine posts fault probability and FFT spectra per measurement window
Your own gatewayPosts readings to the OpexMX API with a connector API key
Manual entryOperator rounds and meter readings typed in the web or Field app

Setup, payload formats and limits are on Integrate with Machines. Every reading passes a data-quality check before it is stored (see Data Quality below).

Health Score

Each asset gets a score from 0 to 100, so the machines that need attention sort to the top. The score combines these factors:

FactorBase weightWhat it looks at
Parameter degradation40%How far current readings have drifted from each parameter's baseline
Failure history25%Failures in the last 30 and 90 days
PM compliance20%Overdue preventive maintenance
Asset age15%How old the asset is; the score falls off around 10 years
Vibration health25%ISO 20816 zone and model fault probability, only for assets with a vibration sensor

Weights are rescaled over the factors an asset actually has, so an asset without a vibration sensor is scored on the first four alone. The asset page shows each factor's score and weight.

Scores are recalculated every day at 06:00 UTC, within minutes of new machine data (at most once per 10 minutes per asset), and immediately when a vibration alert opens. Recalculate All forces a full run. It sits on the predictive maintenance dashboard (/predictive/maintenance in your OpexMX address), which also holds the Anomalies, Predictions, Alarms, Regimes, RUL, Baselines, Recommendations, Cost Model and Part Supply tabs.

Alerts and the Alerts Inbox

Several detectors run on the same data, and their findings all land in the Alerts inbox:

Source in the inboxWhat raises it
Vibration (SE-CNN)The vibration model reports a fault for a run of consecutive windows (3 by default, set per sensor)
Vibration (FFT)The spectrum reaches ISO 20816 zone D, or shows a flagged fault frequency or zone C for the same run of consecutive windows
Threshold crossingA parameter's trend is projected to cross its limit
Multivariate (PCA)Parameters stop moving together the way they normally do for that asset type, even if each one is within its own limit
Model driftThe inputs to a PCA model have shifted over the last 7 days

Predictive alerts are grouped into alarm incidents: duplicates for the same asset and alert type on the same day are merged, each asset raises at most 5 new incidents a day by default, and alerts during a planned PM are shelved instead of paging anyone.

From the inbox you can acknowledge or resolve an alert and create or open its ticket; the Alarms tab of the predictive maintenance dashboard can also shelve or escalate an incident. Resolving asks for a disposition: true positive, false positive, degraded, or sensor issue. True positives become training labels; false positives and sensor issues are kept out.

Vibration sensors create a ticket automatically when they alert (can be turned off per sensor). If the asset already has an open ticket, the alert is linked to it instead of opening another.

For fixed limits that should open a ticket on their own ("bearing temperature above 85 °C for 10 minutes"), use a parameter rule in Ticketing ticket rules: comparison operator (>, <, >=, <=, =, !=), threshold, an optional duration the condition must hold, and a cooldown. The same limit is used as the threshold in Trend Prediction.

Trend Prediction

Trend Prediction answers "when will this parameter cross its limit?" for every monitored signal, and sorts them into Critical, Watch and Nominal (72 hours or more away, or idle).

Five statistical methods are available, each suited to a different pattern:

MethodSuited to
Linear regressionSteady wear, such as a slowly rising bearing temperature
Holt's double exponentialDegradation that is speeding up
EMA with rate of changeNoisy signals
Bollinger bandsEarly warning when a value leaves its normal band
Holt-WintersSeasonal or cyclic patterns, such as shift-driven load

Forecasts use daily averages. A signal with fewer than 14 days of history is marked as insufficient and does not raise an alert, and a forecast never looks further ahead than the length of history behind it. Without a limit set in a ticket rule, the trend is still drawn but no time-to-threshold is given.

Remaining Useful Life

RUL is estimated two ways:

  • Weibull, fitted per asset class from failure history, with completed PMs counted as suspensions (units that were serviced before they failed). Pooling a class lets a small fleet get a fit at all, so treat the class figure as indicative.
  • Trend, from the threshold-crossing projection of a single asset's parameters.

When the running-state gate is enforced, ages are counted in running hours rather than calendar hours, then converted back to days using the asset's recent utilisation. The Weibull and proportional-hazards math comes from @opexcg/reliability, our open-source (Apache-2.0) reliability library.

Asset Condition Page

Click any asset in the Overview or inbox to see, on one page: vibration sensors with the model version they run and the latest FFT spectrum, the top contributing parameters of an active PCA anomaly, the operating-regime timeline (idle to peak load, with startups), RUL, health score factors, open tickets and failure-label history.

Recommendations and Cost Model

For each asset with an RUL estimate, OpexMX scores three actions (planned PM, replace, keep monitoring) by expected total cost: downtime, labour, spares and expedite fees, weighted by the probability of failure. The top recommendation per asset is listed on the Recommendations tab of the predictive maintenance dashboard, and a work order can be created from it.

Costs come from a global cost model (downtime cost per hour, labour rate, expedite fee, spares surcharge, default PM and failure downtime, rate uncertainty) with per-asset overrides, and Part Supply records on-hand quantity, unit cost and vendor lead time for the parts each repair needs.

Outcomes: Did the Prediction Come True?

Outcomes lists completed corrective tickets waiting for an ISO 14224 failure label (failure mode and related coding). These labels are the ground truth for:

  • RUL and Weibull failure history
  • Training data for the vibration model
  • Backtests that measure detector accuracy

A few minutes of labelling per week is what makes the rest of this page trustworthy.

Models & Evaluation

TabWhat it does
ML modelsVibration (SE-CNN) model versions, training runs and datasets; activate a candidate model or retire one
PCA modelsMultivariate models per asset type and operating regime, retrained daily; Retrain and Scan now run them on demand
BacktestReplays the threshold and PCA detectors over labelled failures month by month, training only on data from before each month, and reports precision, recall and median warning lead time. Run on demand; export as PDF or CSV
TuningSuggested threshold changes based on the latest backtest, to apply or reject

Pilot Selection

Starting predictive maintenance on every asset at once rarely works. Pilot Selection ranks candidates by criticality matrix, bad-actor history (last 30 days to last year) and health score, shows a bad-actor Pareto chart, and lets you keep a curated pilot set. Export the list as CSV.

Data Quality

Every reading goes through a signal check before it is stored. Readings that fail are quarantined with a reason instead of being silently dropped or coerced, for example unknown_tag, unit_mismatch, out_of_range, timestamp_future, text_in_numeric or bad_quality.

TabUse it to
OverviewPer-machine data-quality report
QuarantineReview refused readings, then release or discard them
RolloutSwitch each ingest path between off, shadow (log only) and enforce
Running stateCheck the running/idle gate and conflicts between status and signals
Feature storeDaily asset feature rows, build and backfill
EnginesSee why an engine is idle for an asset (for example "not enough history")

New tenants start with every path on enforce. Manual entries that fail the check are refused on the spot with the reason, so the person typing can correct them.

What Runs Automatically

A daily job (06:00 UTC) runs the full chain: running-state gate, health scores, anomaly detection (z-score with slow-drift detection), degradation curves and alerts, operating regimes, PCA training and scan, RUL, alarm consolidation, model drift, baseline comparison, cost recommendations, threshold suggestions, ML dataset and retraining triggers, and stale-sensor checks. Each step records its own status, so a step that failed or had nothing to do shows up on the Engines tab instead of only in a log.

Who Can Do What

PermissionAllows
predictive.readOpen the Condition Monitoring pages
predictive.alert.manageAcknowledge, resolve, shelve and escalate alerts and incidents
predictive.trigger.manageRun scans and model training, manage vibration sensors and ML models, change data-quality rollout and quarantine
predictive.health.recalculateRun Recalculate All on health scores
predictive.pilot.manageEdit the criticality matrix and the pilot set
predictive.iot.ingestSend machine data through the API (held by the user who owns the connector API key)

By default only Admins have the full set. Supervisors and Engineers have predictive.iot.ingest so they can own connector keys. Grant predictive.read and the others per role in User Management.

Getting Started

  1. Pick 5 to 10 pilot assets on Pilot Selection: critical, failure-prone, already instrumented
  2. Connect their data through OpexFlow, the vibration edge or the API, and map tags to asset parameters
  3. Set limits for the key parameters as ticket rules, starting from OEM values
  4. Check Data Quality after the first day and fix units and unmapped tags
  5. Work the Alerts inbox daily and resolve every alert with an honest disposition
  6. Label completed corrective tickets on Outcomes each week
  7. Run a backtest after a few months of labels and apply the tuning suggestions you agree with

Tips

  • A rule duration plus a cooldown cuts most false alarms from brief spikes, such as a startup surge
  • Trend forecasts need at least 14 days of history; don't judge a new signal on its first week
  • A falling health score with no alert usually means overdue PMs or recent failures: open the factor breakdown before sending someone to the machine
  • Mark sensor faults as Sensor issue, not False positive, so model accuracy figures stay meaningful

For AI Agents

Agents connected through the MCP server can read health scores, anomalies, failure predictions, RUL, regimes, alarm incidents, vibration alerts and backtest summaries, and ask for a plain-language explanation of an alert (explain_alert).


Related