False Positives in Predictive Maintenance: When Your Model Is Just Hearing Electrical Noise

Predictive maintenance programs promise fewer surprise failures and lower repair costs. Vibration sensors, current sensors, and temperature probes feed continuous streams of data into machine learning models trained to spot the earliest signs of trouble. A model is only as reliable as the signal it receives, though, and on a busy plant floor that signal picks up more than mechanical wear.

Maintenance teams are increasingly finding that some of their most confident anomaly alerts trace back to something other than a failing bearing or a misaligned shaft. The culprit is electrical noise, and it can look remarkably similar to the early warning signs everyone is watching for.

The Anomaly That Wasn’t There

A current sensor watching a motor’s draw doesn’t know the difference between a genuine mechanical fault and a voltage transient. Both show up as a sudden spike outside the model’s expected range. An anomaly detection algorithm trained to flag deviations will treat them the same way, generating a work order for an inspection that turns up nothing wrong.

This happens more often than most teams expect. Every motor start, every VFD switching event, and every contactor closing nearby sends a brief electrical transient through the wiring. A single false alarm is a minor inconvenience on its own. Across a plant with hundreds of monitored assets, the pattern erodes trust in the entire system.

Where Electrical Noise Comes From on the Plant Floor

Industrial environments generate transients constantly, and most of them have nothing to do with equipment health. Understanding the common sources helps maintenance and data teams separate signal from noise before a model ever sees the data.

• Motor starts and stops, which draw inrush current well above steady-state levels
• Variable frequency drive (VFD) switching, which chops the waveform thousands of times per second
• Contactors and relays opening or closing on nearby circuits
• Inductive loads switching off, which release stored energy as a voltage spike
• Poor grounding or shared wiring runs that let noise couple between circuits

Why Noise Mimics Real Fault Signatures

Bearing wear and shaft looseness both show up as increased vibration and irregular current draw, the same features a transient produces for a fraction of a second. Without careful feature engineering, a model built to catch anomalies will treat a two-millisecond voltage spike from a motor start the same way it treats the slow current drift of a bearing wearing down over weeks.

The distinction matters because the two conditions call for completely different responses. A bearing showing early wear needs a scheduled inspection. A transient needs nothing at all, and chasing it wastes technician hours that should go toward assets with real problems.

Cleaning the Signal Before the Model Ever Sees It

The most effective fix often happens at the hardware level, before the data ever reaches a model. Sensor input circuits that lack basic protection pass every transient straight through as if it were meaningful data.

Without a TVS diode protecting the current sensor’s input stage, every motor start injects a transient that a poorly tuned anomaly detection model can mistake for the onset of bearing failure. Clamping the voltage at the source keeps that spike from ever reaching the analog-to-digital converter, which means the model never has to learn to ignore it in the first place.

• Add transient voltage suppression at every sensor input, not just the ones that have already caused problems
• Route signal wiring away from motor leads and VFD output cables wherever possible
• Use shielded, twisted-pair cable for analog sensor lines and ground the shield at one end only
• Apply low-pass filtering in the data acquisition hardware to remove high-frequency spikes before sampling
• Log known switching events, such as motor starts and drive ramps, so they can be excluded from training data

Building Models That Know the Difference

Hardware fixes reduce the noise reaching the model, but a well-built model should also be able to tell the two conditions apart. Feature engineering that looks at duration and recurrence, not just amplitude, catches this quickly. A transient lasts milliseconds and disappears. A developing bearing fault persists and typically worsens over days or weeks. 

Labeling matters just as much as feature selection. Training data that includes timestamps for known switching events lets a model learn to discount them automatically, rather than flagging every one as a new anomaly. Teams that maintain this kind of event log alongside their sensor data tend to see nuisance alerts drop off within the first few months of tuning.

The Cost of Chasing the Wrong Signal

False alarms carry a real cost even when no equipment is actually at risk. Technicians who investigate the same nuisance alert several times a month start to trust the system less, and that erosion of trust is often what kills a predictive maintenance program long before the technology itself fails.

Getting ahead of electrical noise, in the wiring and in the model, protects the credibility of the entire program. A system that only alerts when something is truly wrong is the one maintenance teams will actually act on.