Failure risk model
A supervised model trained on historical sensor streams and CMMS outcomes that scores each asset's probability of failure within an actionable horizon.
Part of the recommended approachTechnique: Predictive Modelling on Sensor Data
The failure risk model is the ranking layer of unplanned-downtime prevention, taking the assets that condition threshold alarms and sensor drift screening surface and ordering the whole monitored population by predicted probability of failure within a defined horizon.
It looks at trends across vibration RMS, spindle motor current, bearing temperature, and oil particulate counts together, rather than any one reading against a fixed set point, so it can catch degradation that only shows up as a joint pattern.
Its output is a risk score with an estimated horizon and the specific readings and trend that drove it, handed to a maintenance planner deciding what to inspect during the next planned maintenance window.
This is the wrong tool for cheap, redundant, non-critical components already on a run-to-failure strategy, since a prediction there changes no decision and only adds noise to the queue.
It is also the wrong tool wherever confirmed failure history is too thin or too distorted by survivorship bias to train against honestly, since a model built on that history will reproduce the plant’s existing maintenance schedule rather than learn anything new about equipment condition.
Tune the score toward precision rather than recall for any asset where the response is an expensive teardown, because a false alarm there creates a failure risk of its own on a component that was healthy.
Revalidate the model whenever instrumentation changes, since a silent shift in the input distribution is a common way this kind of model fails in production.
Strengths for this problem
- Combines weak signals across vibration RMS, spindle motor current, and bearing temperature into a single ranked score, catching combinations no individual threshold would flag.
- Ranks the entire monitored asset population, so a planner works the highest-risk equipment first instead of the queue order alarms happen to produce.
- Can extend effective lead time beyond what a single threshold gives, when trained on a trend rather than a single crossed value.
Limitations for this problem
- Needs a meaningful volume of confirmed failure events, corrected for survivorship bias, or it learns the existing maintenance schedule instead of actual equipment health.
- Degrades quietly when sensor calibration, sensor replacement, or setpoint changes shift the input distribution without anyone flagging it.
- A false positive drives an unnecessary teardown, which introduces its own failure risk, so the model must be tuned for precision on expensive interventions rather than for recall.
Data needed
- Historical vibration RMS, spindle motor current, bearing temperature, and oil particulate counts, joined to confirmed failure and repair events by asset ID.
- CMMS work order history with parts replaced and open and close dates, used as the outcome label.
- Asset criticality and redundancy status, to decide which score tier routes to an immediate response versus the next planned window.
- Planned maintenance window calendar, to check that a predicted horizon leaves enough lead time to act.
How it is measured
- Precision of high-risk scores at the lead time actually needed to act within a planned maintenance window.
- Unplanned downtime hours avoided per asset class, against the baseline before the model.
- Rate of unnecessary teardowns triggered by a false positive.
- Score drift, tracked against known sensor recalibration, replacement, and setpoint-change events.
Pitfalls
- Training on raw CMMS history without correcting for survivorship bias, so the model learns the maintenance schedule the plant already follows rather than actual equipment health.
- Treating a work order's close date as the moment the fault began, which mislabels the true failure onset and corrupts the lead-time estimate.
- Continuing to trust a score after a sensor swap or setpoint change without retraining or revalidating it.
- Optimising for recall on a score that feeds an expensive teardown decision, which drives false alarms and technician distrust.
Human oversight
A maintenance planner or reliability engineer reviews every high-risk score and decides whether and when to act. The model never schedules a teardown, orders parts, or stops a line on its own, and any coordinated shutdown is confirmed by a reliability engineer first.