Predictive maintenance starts with signal quality

Predictive maintenance projects often begin with a model and end with an alarm nobody trusts. The failure usually starts earlier: a vibration sensor has the wrong bandwidth, timestamps drift, maintenance records lack failure modes, or the model sees a startup transient as a bearing defect. Signal quality is the product before prediction is the product.

Match the sensor to the fault

Define the failure mode and its physical signature first. A slow temperature trend, a high-frequency bearing fault, a motor-current imbalance, and a lubrication problem do not demand the same sensor or sample rate. Check mounting, orientation, range, noise floor, sampling interval, and missing-data behavior. A dashboard that interpolates a gap can look smooth while erasing the evidence of an outage.

Preserve timestamps at the source and carry quality through every transport. Align measurements with speed, load, recipe, ambient conditions, and operating mode. A vibration spike during a known acceleration may be normal; the same spike at steady state may deserve inspection.

Labels are operational decisions

Failure labels should describe what happened, when it became detectable, what work was performed, and whether the work fixed the cause. “Bearing replaced” is not enough if the root cause was misalignment or a loose mount. Ask technicians for a small, consistent vocabulary and allow free text as evidence rather than as the only label.

Start with a baseline: threshold, trend rule, or spectral band. Compare any machine-learning model against it on the same time window. Track false alarms per asset-month, detection lead time, missed failures, and the percentage of alerts that become useful work—not only precision in a notebook.

Close the maintenance loop

Send an alert with the evidence and recommended next step to the CMMS or maintenance workflow, with a human approval step. When a technician dismisses or confirms an alert, capture that outcome. Review performance by asset class and operating mode; a model that works on one pump family may be wrong for another.

The predictive maintenance data stack describes the layers from sensing to work order. The condition monitoring glossary entry adds the vocabulary. For the wider architecture, Understanding the Shop Floor is a practical ebook companion.