The maintenance is predictive. The process change is invisible.

The vibration sensor gave the spindle six weeks to live. Maintenance planned the swap for a Saturday, executed it clean, and booked the avoided breakdown on the monthly review as a €38,000 win. Nobody in that room knew the surface finish on the customer's sealing face had drifted out of spec four weeks earlier – inside the exact window the bearing had been telling its own story. The customer knew. There was a return in the containment cage with a rejection letter attached. That is what predictive maintenance predicts. And what it doesn't.

Two clocks, one machine

Every machine runs two clocks. The breakdown clock – bearings, spindles, drives, ball screws – is physics grinding toward failure, and vibration analysis reads it well. The capability clock runs faster. Runout, backlash, thermal growth, chatter – the same degradation that will eventually stop the machine starts bending its geometry long before any health threshold trips. I have watched a forming line sit at a stable Cpk 1.5, then slide to 0.9 across three weeks while the condition-monitoring dashboard held a confident green. Both readings were honest. They were answering different questions.

The uncomfortable part: the clocks are not independent. The vibration signature that predicts a spindle's death is the same vibration degrading your surface finish today. A machine in the final weeks of component life is, by definition, a process whose inputs are drifting. The health dashboard says "no action required – failure predicted in 42 days." The control chart says special cause. Only one of those reports reaches your customer.

Uptime is a maintenance metric; conformance is a process metric. Confuse the two and you will ship beautiful scrap on schedule.

Every predictive fix is an unannounced process change

Here is the part the vendor demos skip. Predictive maintenance does not remove interventions. It reschedules them. And every intervention is a process change wearing overalls.

At WITTE Automotive I spent months in QRQC sessions chasing a dimensional drift that arrived, vanished, and arrived again. The parts held no pattern. The nights did. The drift tracked to overnight maintenance nobody had logged: a re-tensioned drive on Tuesday, a replaced hydraulic cylinder on Friday, a 02:00 component swap done to protect the morning schedule. First-article inspection never ran, because – in the shift's words – it was only maintenance. Only maintenance. As though a new spindle arrived carrying the old process dynamics. It carries new runout, new thermal behaviour – in effect a different machine from the one your control plan was validated against.

The A3 that closed it held no exotic countermeasure. It held a rule: any work order that touches the process is a change under the QMS and triggers the reaction plan, same as a tooling change – because that is what it is. A large share of the failure-cost reduction we delivered there came not from better measurement but from making that class of silent change audible.

What the model cannot see

The trade press is busy selling predictive maintenance as risk minimisation again. On the breakdown clock, fair enough – avoided unplanned stops pay for the sensors, and I will not argue with that arithmetic.

But I also build multi-model AI systems – orchestration, not model training – and that work teaches you the edge of a model's perception first. A predictive-maintenance model ingests vibration, temperature, current draw. It has no idea what the control chart knows: the mean shift that began the morning after the spindle swap, the tightening spread on a bore, the special-cause signature no bearing signal carries. My own platform runs 63 models to consensus before answering anything, and I would still not ask it about conformance. Consensus among models that cannot see a domain is agreement, not insight. Crossing that boundary is not a modelling problem. It is a governance problem.

Wiring maintenance into the quality system

When I built the quality function at SNOP's greenfield plant – 900+ employees, QA/QC from nothing – this wiring was one of the fights worth having. Maintenance owned the CMMS. Quality owned the control plan. Neither system knew the other existed, and neither team was eager to share. The fix was procedural before it was technical: a work order flagged as process-touching fires a quality notification automatically, and parts run after the intervention are verified against the control plan – first-piece as the floor, a short capability check on critical characteristics where risk justifies it. Maintenance keeps its 02:00 authority. The quality engineer reads about it at 07:20, before the day's first pallet ships.

The 70% defect-cost reduction that plant delivered had several fathers. One was this: defects stopped arriving with mysterious causes, because the causes were no longer silent. And the vendors are not waiting for your QMS to feel ready – component makers are certifying to IATF 16949 specifically to sell the instrumented plant into automotive supply chains. The dashboards land whether quality is wired in or not.

Buy the sensors. The economics are real and the avoided downtime is real. But be clear about what the purchase delivers. Not uptime – a handshake: when maintenance touches the process, quality requalifies it. Predictive maintenance without that handshake does not minimise your risk. It schedules your scrap more efficiently.