Three Approaches to Predictive Maintenance in AI Data Centers
Predictive maintenance covers three distinct methods: schedule-based, condition-based, and physics-based early warning grounded in waveform physics. Most vendors offer the first two. Here is how they compare.
I was at a customer site last month looking at a fleet of UPS rectifiers. Every unit was inside its maintenance window. Every one had passed its most recent scheduled inspection. The site team had done everything by the book.
Four percent of them were actively degrading. We caught it through waveform analysis at 8 kHz. No alarm had fired. No threshold had been crossed. The maintenance records said "healthy." The physics said otherwise.
This keeps happening because "predictive maintenance" has become a label that covers three fundamentally different approaches. In our world, predictive maintenance means early warning grounded in waveform physics. Schneider uses the term differently. Eaton uses it. Vertiv, ABB, Siemens. They all use it. The approaches they describe predict at different timescales, detect different failure modes, and require different instrumentation.
If you are evaluating vendors right now, the label will not help you. The method will. Verdigris's approach to predictive maintenance is early warning grounded in waveform physics: what changes inside the equipment, captured at 8 kHz, before any threshold is crossed.
4%
Degrading rectifiers caught between scheduled inspections
21 days
Validated lead time before failure with physics-based detection
3× nameplate
Capacitor ESR on units a scheduled inspection had cleared two weeks earlier
Three approaches, three detection horizons
Schedule-based: the actuarial model
This is prediction in the statistical sense. The manufacturer says replace UPS capacitors every five years. Your CMMS generates the work order. The interval is based on population-average failure rates, not on what is happening inside any individual unit.
The 5-year-old capacitor that is performing fine? Gets replaced. The 2-year-old capacitor that thermal cycling has already degraded? Not on the schedule for another three years.
Schedule-based programs protect you from the average failure. They do not protect you from the outlier. And in my experience, it is the outlier that causes the outage.
Condition-based: the threshold model
Condition-based monitoring improves on schedules by looking at the equipment itself. Temperature sensors, vibration probes, oil analysis. When a reading crosses a threshold, you take action.
The limitation is straightforward: you can only detect degradation in the parameters you are monitoring. For UPS rectifiers, the critical degradation shows up in the electrical waveform, not in temperature or vibration. By the time thermal sensors pick up rectifier capacitor degradation, you are often weeks past the point where you had comfortable lead time.
I have seen condition-based programs that monitored eight parameters per UPS and still missed the failure mode that actually occurred. Not because the system failed. Because the failure mode expressed itself in a parameter the system was not measuring.
The blind-spot in condition-based monitoring
You can only detect degradation in the parameters you are monitoring. Eight monitored parameters per UPS is not a small instrumentation footprint. It still misses electrical failure modes that don't express themselves as temperature, vibration, or oil-chemistry shifts.
Physics-based: the waveform model
Physics-based early warning works differently. Instead of checking specific parameters against thresholds, it captures the full electrical behavior of the equipment and watches for changes in the underlying physics.
For electrical infrastructure, that means 8 kHz waveform capture that resolves harmonic content through the 25th order. When a UPS rectifier capacitor starts losing capacitance, it shows up as elevated 5th and 7th harmonic THD. That shift is detectable weeks before the rectifier drops below operational thresholds.
Each failure mode has a distinct electrical signature. Capacitor loss looks different from switching device wear, which looks different from transformer winding stress. The waveform tells you what is degrading, not only that something changed.
Comparing the three
When each approach fires, relative to a UPS rectifier failure
Each row marks where on the timeline an approach typically alerts you. Earlier (left) means more lead time. Wider band means less precision about when the alert fires.
Physics-basedH5 / H7 harmonic ratios shift 21 days before the rectifier crosses operational thresholds. Per-unit, not fleet-level.
The three approaches predict at different timescales. More importantly, they answer different questions. Schedule-based predicts the population. Condition-based predicts the threshold. Physics-based predicts the individual unit.
The diagram shows when each approach detects. The table below shows what each approach is actually for.
| Aspect | Schedule-based | Condition-based | Physics-based |
|---|---|---|---|
| Replaces by | Calendar interval | Threshold crossing | Individual degradation signature |
| Failure modes covered | Predictable consumable wear | Mechanical (bearings, oil, temperature) | Electrical (capacitor ESR drift, harmonic shifts) |
| Cost of a false positive | Built-in (healthy units replaced on schedule) | Tuning trade-off between sensitivity and noise | Low (physics-grounded anomalies) |
| Instrumentation footprint | None (calendar-based) | Per-sensor wiring plus threshold setup | 8 kHz waveform capture per circuit |
| Where it shines | Consumables with predictable wear | Mechanical degradation | Electrical degradation invisible to BMS |
In a large fleet, the difference between replacing by cohort and replacing by individual signature can be servicing a handful of units instead of hundreds.
What this looks like in the field
Different site, last year. The ops team had completed a full scheduled maintenance cycle on their UPS fleet the prior month. Every unit inspected. Every checklist completed. Clean bill of health.
Two weeks later, our waveform analysis flagged three rectifiers with elevated 5th and 7th harmonic ratios consistent with capacitor ESR increase. The maintenance team was skeptical, since they had inspected those units the week before. We pulled the capacitors. ESR was 3× nameplate on two of them, 2.4× on the third. The scheduled inspection had checked the units visually and verified operational parameters. The degradation was electrical, not visible, and the operational parameters were still within 5% of spec.
Three rectifiers, two weeks after a clean inspection. The schedule didn't catch it because the schedule wasn't looking for it.
What I keep coming back to is that the inspection wasn't bad work. The team did everything the schedule asked. The schedule was the problem.
That is the difference I try to get across when I am talking to ops teams. It is not about having more data or better dashboards. It is about measuring at the right timescale, in the right domain, for the failure mode that actually threatens your uptime.
Why this gets harder with GPU loads
AI data centers make this conversation more urgent. GPU-dense racks create harmonic profiles that stress electrical infrastructure differently from CPU workloads. The degradation timeline changes. Schedule-based intervals calibrated for 5 kW server racks may be too long or too short for a 120 kW training rack. You do not know which one until you measure.
The cost side is different too. A UPS failure that interrupts a multi-week training job on hundreds of GPUs is a different financial event than a UPS failure that bounces a fleet of web servers. The training job may require days to restart from the last checkpoint. You cannot reboot and carry on.
And the resolution side compounds it. 1 Hz polling cannot see the sub-cycle electrical changes that precede rectifier failure. That is the same sampling gap that makes training jobs die without a useful error in the logs.
What I would ask your vendor
If you are evaluating predictive maintenance offerings, you will see two kinds of vendor in the room: those promising early warning grounded in waveform physics, and those promising nicer dashboards. Here are the questions I would bring to either:
- What is the detection basis? Calendar age, sensor thresholds, or waveform physics resolved through the 25th harmonic order?
- What is the minimum lead time before failure, and what determines it? Ask for specific examples from deployed sites (21-day, 7-day, day-of-failure), not theoretical numbers.
- What instrumentation is required? If the answer is "existing BMS data," sanity check what failure modes that data can actually detect. A 1 Hz temperature poll will not catch harmonic degradation that lives between samples.
- Can you show per-unit identification, or does the system flag at the fleet level? Knowing that "something in your UPS fleet is degrading" is not the same as knowing which of your 800 rectifiers needs attention.
- What is the false positive rate in production? Schedule-based approaches replace healthy equipment by design. That is a false positive in everything but name.
The right answer may be a combination of approaches. Schedule-based programs still make sense for consumables. Condition-based monitoring covers mechanical failure modes well. But for the electrical infrastructure that stands between your utility feed and your compute, physics-based early warning provides detection that the other approaches cannot.
For UPS failure prediction using physics-based early warning, see the dedicated solution page.
Talk to an engineer about adding physics-based early warning to your maintenance program.