On a plant floor there are two ways for a warning system to be wrong. A missed failure means unplanned downtime, secondary damage, a blown schedule and, at the limit, a safety incident. A false alarm means a wasted inspection. That is tolerable once. Repeated, it teaches the crew to ignore the system, and then a real warning arrives to people who have stopped listening. The second failure is the one that kills these programmes, because it is gradual and nobody records the day the system stopped being believed.
A machine usually announces a failure long before it fails, and predictive maintenance is the work of hearing it in time.
Most mechanical failures are not instantaneous. A bearing begins to wear, a shaft drifts out of alignment, a winding starts to run hot. Long before the machine stops, those changes show up as small shifts in vibration, motor current or temperature. Reliability engineers describe this with the P-F interval: the stretch of time between the point where a developing failure first becomes detectable and the point where the machine actually fails. Predictive maintenance lives inside that interval. A warning near its start leaves time to order parts and plan the work. A warning near its end is barely better than the failure itself.
Detecting the onset sounds like a classification problem, and that framing causes trouble. Failures are rare, so almost every reading is healthy, and a system that never raises a warning scores very well on accuracy while catching nothing. The useful way to think about it is as two error rates pulling against each other. Lower the alert threshold and you catch more real failures but raise more false alarms. Raise it and the false alarms fall but more failures slip through. Where to sit on that trade-off depends entirely on what each kind of error costs the plant, and those costs are almost never equal.
There is also the question of what normal means. A pump running hard under heavy load vibrates differently from the same healthy pump idling. If every reading is compared with a single average, each change of load or speed looks like an anomaly. Comparing a reading with what is normal for the current operating condition removes that whole class of false alarm, and it requires knowing how the equipment behaves as much as how the data looks.
- Machine running healthyreadings sit within normal for this load and speed
- Failure begins to developwear, misalignment or heat, not yet visible
- First detectable changethe P point, where the window opens
- Condition indicator trendsvibration or current moving steadily
- Warning with lead timeearly enough to plan the work
- Functional failurethe F point, where the window closes
Accuracy is the wrong target by construction. Failures are rare, so a model that always reports healthy scores over ninety-nine percent and catches nothing. Two more problems sit underneath. The same healthy machine produces different vibration and current signatures at different loads and speeds, so a reading judged against one global baseline flags normal operation as anomalous every time the operating regime changes, and that is the dominant source of false alarms. And sensor data from a working plant is not clean. A dropout or a slow calibration drift can mimic a fault convincingly if nothing upstream conditions the signal.
Predictive maintenance usually fails by being disbelieved, long before it fails by being wrong.
The common story is not a system that misses a dramatic failure. It is a system that raises too many alarms in its first weeks, each one checked and found to be nothing, until the crew stops checking. From then on the system is still running, and still correct some of the time, but nobody acts on it. When the real warning arrives, it is treated like all the others.
The causes of those false alarms usually sit upstream of the model. Readings judged against the wrong baseline, sensors that drop out or drift, and thresholds chosen to maximise a score all produce alerts that look like detections. A model can be correct about the data it was given and still be wrong about the machine, because the data was never conditioned or put into context.
The reflex, and the fix.
Tune for accuracy, alert on anomalies
Standard, quick, and it produces a high score on historical data.
It either misses the rare failures or floods the crew with false alarms, and either way the system is switched off within a month.
Cost-weighted, with actionable lead time
Normalisation for operating context, a threshold tuned to cost, a lead-time gate, and calibration against real run-to-failure history. More to build and more to tune.
Warnings the crew believes, early enough to act on.
Optimise for the cost of the plant's errors and the timeliness of its warnings, not for a metric on a slide.
Condition the signal first
Denoise, fill and resample gaps, remove drift and align sensors to a clean baseline before anything else runs. Conditioning removes false signals that no downstream model can recover from. Half of what looks like a detection problem is a signal processing problem nobody addressed.
Normalise for operating context
Each reading is judged against the baseline for the current load and speed, not against one global baseline. This single decision removes the largest source of false alarms, and it depends on understanding the equipment as much as the dataset.
Predict only what leaves a trail
The system tracks condition indicators that trend as a machine degrades, and predicts the crossing before it happens. Part of its honesty is knowing which failure modes degrade gradually and can be seen coming, and which arrive suddenly and cannot.
A decision layer, not a score
An alert fires only when a cost-weighted threshold and a lead-time gate both agree. Both are calibrated on real run-to-failure history and held just short of the false-alarm rate that would cost the crew's trust. Everything else stays quiet, on purpose.
What each choice cost, and what it bought.
ATuning to cost, not to accuracy
A threshold chosen to maximise accuracy is easy to justify on a slide. On a plant floor it is meaningless, because a missed failure and a false alarm cost wildly different amounts. Setting the operating point by the real cost of each error means the plant's economics have to be understood, and the threshold has to be revisited as they change. It buys an alerting system aimed at the plant's actual problem.
BConditioning and context before modelling
It is tempting to start with the model, because that is the visible and interesting part. Conditioning the signal and normalising for operating context first delays any model result. It also removes the largest sources of false alarms before a model ever sees them, which no amount of downstream modelling could have done.
CRequiring both a cost threshold and a lead-time gate
A single score crossing a line is the simplest possible alert. Requiring two conditions to agree, that the risk is worth acting on and that there is still time to act, means some warnings are held back. The ones that remain are warnings the crew can use, and every alert they act on successfully makes the next one easier to believe.
DBeing explicit about what cannot be predicted
Claiming coverage of every failure would have made the system sound more capable. Saying plainly that sudden and random failures leave no trail narrows the promise. It also means the crew knows which failures the system is watching for and which still need other safeguards, so nothing is assumed to be covered when it is not.
A single headline number hides where a system fails. This work was scored on the dimensions that actually decide whether it holds in production, measured on real, held-out cases rather than the demo path.
A high score on a slide and, on the plant floor, either silence on the failures that matter or a stream of false alarms that gets the system muted within a month.
False alarms fell 41% across the 87-unit fleet, and the alerts that remained were believed and acted on, early enough to prevent the failures that mattered. The operating point was set by the real cost of each kind of error and not by a generic accuracy figure.
What it owns, and what it hands to a person.
It predicts the degradation modes it has signal and history for. A truly sudden or random failure leaves nothing to warn on, and no system will change that. Conditions outside its training degrade it. The cost threshold has to be maintained as the economics and the equipment change. And it informs a maintenance decision. It does not make one.
Three of the five decisions that carried this work sit below the model: signal conditioning, operating-context normalisation and the choice of health indicators. A specialist hired only to build the model would have built a correct model on an uncorrected signal, judged against the wrong baseline, and would have been entirely within scope while the system failed. The discipline carries to any detection problem with rare, expensive events. Weight the errors by what they cost, normalise for conditions, demand actionable lead time, calibrate for trust, and be honest about what has no signal at all.
When failures are rare, accuracy is a misleading target. A system that never alerts can score very highly and catch nothing.
Set an alert threshold by what each kind of error costs, because a missed failure and a false alarm are almost never equally expensive.
Compare readings with what is normal for the current operating condition, or every change of load will look like a fault.
A warning is only useful if it arrives with enough lead time to act on. Measure lead time as carefully as detection.
The quickest way to kill a monitoring system is a stream of false alarms. Trust is lost gradually and is very hard to win back.