On a CNC, only some failures announce themselves.
Which components give real early warning, which never will, and why per-machine baselines are the difference between an alert that means something and one that gets muted.
Predictive maintenance on a machining centre is a narrower problem than the phrase suggests, and being precise about it is the difference between a system that earns trust and one that gets muted after the third false alarm.
A CNC has a small number of components whose failure is both expensive and preceded by measurable change: the spindle bearings, the ballscrews and their bearings, the axis drives, the tool changer, and the coolant and hydraulic systems around them. Most of what actually stops a machining centre — a crashed tool, a program error, a fixture problem — is not predictable from vibration and never will be.
Component by component.
The high-value case
Defect frequencies appear in the envelope spectrum well before noise or heat. This is the component where early warning is worth the most, because a spindle rebuild costs both the repair and the weeks of waiting.
Slow, measurable degradation
Preload loss and wear show as changing current signatures and vibration during rapid moves. Trend against the machine's own history rather than a threshold.
Mechanical repetition
ATC cycles are highly repeatable, so a change in timing or current draw during a change is a reliable early indicator of a mechanical problem.
Growth and drift
Warm-up curves and steady-state temperatures are stable per machine. A machine taking longer to stabilise, or stabilising higher, is telling you something before dimensional drift reaches the part.
The boundary that vendors blur.
Tool breakage from a bad program, a crash, a fixture that was not clamped, a bar that was out of spec — these are events, not degradation. No amount of vibration data anticipates them, and a system that claims to predict them will produce alarms that are wrong often enough to be ignored, which then destroys trust in the alerts that were real.
The honest framing is that predictive maintenance addresses the wear-out failures, which are a minority of stoppages but a large share of the cost, while downtime capture and OEE address the event failures, which are the majority of stoppages. A plant needs both, and confusing them is why some predictive projects are judged failures despite working exactly as designed.
Events cannot be predicted. Some of them can still be caught.
There is one useful qualification to everything above. A crash or a heavy impact is not predictable — but it is unmistakable the instant it happens, because it produces an acceleration signature nothing in normal machining resembles. High-performance accelerometers sample fast enough to register that event rather than averaging it away, which is the difference between detecting a crash and noticing it later from the scrap.
That matters for two reasons. The first is evidence: an impact that is timestamped and recorded can be investigated, attributed to a program, a job or a shift, and correlated with what the spindle did afterwards. The second is condition: a spindle that has taken a hit often behaves differently from that moment on, and having a marked event in the history makes the subsequent change in its signature interpretable rather than mysterious.
This is part of the vibration monitoring module rather than the predictive layer, and it is worth specifying deliberately if crashes are a real cost on your floor.
Sensing, baselines, and what triggers an alert.
Tri-axial accelerometers are mounted at the spindle housing and, where justified, at axis bearing blocks. Sampling is set for the speed range that machine actually runs rather than a generic configuration, because a spindle running 800–12,000 rev/min moves its defect frequencies constantly — which is why live spindle speed from the control matters as much as the vibration itself.
Each machine learns its own baseline. This is not a nicety: two identical VMCs installed the same week, on different foundations, running different work, have different healthy signatures. Alerting against a handbook limit produces noise on one and silence on the other. Alerting against a slope in that machine's own defect bands produces a warning that means something.
An alert names the component and the evidence — which frequency band is rising, over what period — and converts to a maintenance ticket in one tap. Where the estimate is uncertain, it says so. A prediction with fake precision is worse than a range honestly stated.
What the sequence looks like in practice.
A spindle on a machining centre begins showing rising energy in the outer-race defect band, roughly 3.6× shaft speed at the running speed for that job. Overall vibration is unchanged and nothing is audible. Over the following fortnight the band continues to climb while the rest of the spectrum stays flat, and bearing temperature rises slightly at steady state.
The plant orders the bearing, schedules the change into a planned shutdown, and replaces it during hours that were already lost. The alternative — the same failure discovered when the spindle became audible — is an unplanned stop of unknown duration on a machine that may be the constraint, plus expedited procurement, plus whatever the parts in progress were worth. The bearing frequency calculator shows how that 3.6× figure is derived for your own bearings.