ML
Isolation Forest for sensor anomalies: how it works and where it misleads
Read flagged rows in a sensor table by their isolation score, understand what contamination does in the model, and see where the method misleads you.
Build a predictive maintenance pilot around a decision
Connect sensor behavior to operating context, maintenance events and a false-alarm budget.
In short
- A pilot should first write down the maintenance decision the warning will support, such as scheduling an inspection in the next shift.
- Try a simple robust statistical baseline first and compare any candidate model against it over the same evaluation period.
- Starting in shadow mode records warnings without automatic intervention, and false alerts per asset per day, missed events and lead time are reported together.
Specify the decision
Write down what the maintenance team will do with a warning, such as scheduling an inspection next shift. The warning lead time must allow that action. A model that detects the instant of failure may be too late for planning. Store speed, load and production recipe alongside sensor values.
Link labels to work orders
Align sensor timestamps with work orders, shutdowns and replaced components. Not every vibration increase is a fault: startup and load changes also matter. Unknown labels must not silently become normal labels. Document units, sampling intervals and sensor replacements. Filling missing readings with zero can manufacture a healthy-looking interval.
Establish a simple baseline
Try an operating-regime threshold or a robust statistical baseline first. Kutup Lab demonstrates median and MAD. Compare a candidate such as Isolation Forest over the same evaluation period. An anomaly score is not a failure probability; without verified event labels you cannot calculate fault precision or recall.
Start in shadow mode
Initially record warnings without allowing automatic intervention. Collect the maintenance team’s outcome for every alert. Report false alerts per asset per day, missed events and warning lead time together. Success on one pump does not demonstrate generalization to a different pump or operating regime.
Frequently asked
Where should a predictive maintenance pilot start?
Begin by writing down the decision the maintenance team will make from a warning, such as scheduling an inspection in the next shift. The warning lead time must allow that action. Store speed, load and production recipe with sensor values, and set a simple baseline before trying a more complex model.
How do I measure false alarms in a predictive maintenance pilot?
Start in shadow mode, recording warnings without allowing automatic intervention. Collect the maintenance team's outcome for every alert. Report false alerts per asset per day, missed events and warning lead time together. Success on one pump does not show that the result carries over to a different pump or operating regime.
Sources and further reading
Pilot steps in this guide are proposed engineering workflows. Examples do not constitute measured field performance.
Questions on this topic
What is a good OEE value? Is 85 % realistic?Can equipment faults be detected without labelled failure data?How much data do I need for predictive maintenance?What is predictive maintenance, and how does it differ from preventive?What is remaining useful life (RUL) and how is it estimated?What are MTBF and MTTR, and how are they calculated?
Keep reading.
Full library
ML
Multivariate process monitoring with Hotelling T² and SPE
Read correlated sensors through T² and SPE together rather than one statistic, interpret contribution plots, and learn where the method’s assumptions break.
ML
Without fault labels: unsupervised machine learning in industry
Learn what unlabelled sensor history can and cannot tell you, how to evaluate a detector without labels, and why operating modes need attention first.