# Predictive maintenance on the factory floor
A CNC (Computer Numerical Control) spindle spins at 12,000 RPM cutting aerospace brackets. Three weeks before it would have seized mid-shift, its vibration sensors caught a faint rising signature in a specific frequency band. The maintenance team ordered a bearing, scheduled a two-hour swap for the next planned stop, and never lost production. No fire drill. No scrapped parts. No overtime.
That is predictive maintenance (often shortened to PdM): using machine data to predict failures early enough to fix them on your schedule, not the machine's.
Unplanned downtime is one of the most expensive events on a factory floor. A stopped line can idle dozens of workers, spoil in-process material, and blow through delivery commitments. Industry surveys routinely cite unplanned downtime costs in the tens of thousands of dollars per hour for automotive and heavy manufacturing, though the exact figure varies widely by plant.
Traditional maintenance falls into two buckets, both wasteful:
Predictive maintenance is the third path: intervene based on the actual condition of the equipment, measured continuously.
Each of these is a time series: a sequence of measurements over time. The core insight of PdM is that failures rarely happen instantly. They announce themselves as slow drifts and subtle pattern changes in these streams.
Here is the pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète →, in plain terms.
Sensors sample at a chosen rate (vibration might be sampled thousands of times per second). Raw data is noisy: dropped readings, sensor glitches, a forklift bumping a machine. You filter and align it before anything else.
You rarely feed raw waveforms to a model directly. Instead you compute features: summary numbers that capture the signal's shape. For vibration, engineers use an FFT (Fast Fourier Transform), a math technique that breaks a signal into its component frequencies. A healthy bearing and a worn one look almost identical in raw form but obviously different in the frequency view.
# Simplified: turn a vibration snippet into frequency features
import numpy as np
def bearing_features(signal, sample_rate):
spectrum = np.abs(np.fft.rfft(signal))
freqs = np.fft.rfftfreq(len(signal), 1 / sample_rate)
# Energy in the band where bearing-defect harmonics appear
band = (freqs > 2000) & (freqs < 5000)
defect_energy = spectrum[band].sum()
return {
"rms": np.sqrt(np.mean(signal ** 2)), # overall vibration level
"peak": np.max(np.abs(signal)),
"defect_band_energy": defect_energy, # rises as bearing degrades
}That defect_band_energy climbing week over week is exactly what flagged our opening spindle.
Two model families do the heavy lifting:
A prediction that no one acts on is worthless. The model output must flow into a CMMS (Computerized Maintenance Management System), the software that manages work orders, so a technician gets a ticket, a part gets ordered, and the fix lands on the next planned stop.
Picture a pump on a cooling loop.
1. Baseline: three months of normal vibration and temperature data establish "healthy."
2. Drift: over two weeks, the defect-band energy creeps up 15 percent and the bearing temperature runs slightly hotter each day.
3. Alert: the anomaly model flags it. Severity is "watch," not "urgent."
4. Confirmation: a technician does a quick ultrasound check and confirms early bearing wear.
5. Action: a work order schedules the swap during Sunday's planned maintenance window.
The failure that would have caused an unplanned Tuesday shutdown becomes a routine Sunday task.
Chasing RUL too early. Precise "days until failure" predictions are seductive but data-hungry. Most plants get 80 percent of the value from simple anomaly detection first. Start there.
No failure data. If your machines rarely fail (good news operationally), you have almost no examples to learn from. This is why anomaly detection, which only needs healthy data, dominates real deployments.
Alert fatigue. A model that cries wolf gets ignored. Tune thresholds conservatively and route alerts by severity. A "watch" and an "act now" should look different.
Skipping the humans. Maintenance technicians know these machines intimately. Their knowledge of which failure modes matter and which noises are harmless is training data you cannot buy. Involve them from day one.
Ignoring the last mile. The model is maybe 30 percent of the work. Integrating with the CMMS, spare-parts inventory, and shift schedules is the other 70 percent, and it is what determines whether you actually avoid downtime.
You do not deploy PdM everywhere. You target it. A useful screen: high consequence of failure (safety, big downtime cost, or long lead-time spare parts) plus a measurable failure signature. A critical, hard-to-replace pump is a great candidate. A cheap, redundant fan that fails gracefully is not.
The US Department of Energy's O&M guide is a solid free reference on maintenance strategy tradeoffs, including how condition-based approaches compare on cost.
Start with a pilot on one asset class, prove the savings, then scale. Plants that try to instrument everything at once usually stall under the data and integration load.
Vérification des acquis
1. What is the fundamental principle that distinguishes predictive maintenance from other maintenance strategies?
2. A plant replaces all pump bearings every 3 months regardless of their condition. What is the primary inefficiency of this approach?
3. Why is vibration considered the richest signal for monitoring rotating equipment like spindles and motors?
4. Select ALL correct answers. Which outcomes are characteristic benefits of successful predictive maintenance, as illustrated by the CNC spindle example?
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers. Which statements accurately describe the two traditional maintenance approaches?
Sélectionnez toutes les réponses correctes.
When you pitch PdM internally, frame it in the language operations leaders already use:
Avoid promising a specific percentage improvement upfront. Results depend heavily on your equipment, data qualitydata qualityThe degree to which data is fit for purpose: accurate, complete, consistent, timely, valid and unique. Poor quality data undermines analytics, reporting and AI.Voir la définition complète →, and how disciplined the maintenance response is. Instead, commit to a measured pilot with a clear baseline so the savings are provable, not asserted.
One more practical note for 2026: sensor and edge-computing hardware has gotten cheap enough that retrofitting older machines with vibration and temperature monitoring is often affordable, so you no longer need a brand-new "smart" machine to start.