Predictive maintenance in UPS: how to stop replacing batteries blindly (PdM with AI + thermography)

September 19, 2026 4 min read By Noxtel

Your UPS batteries have been aging since the day they were installed. At 18 months, between 30% and 40% of modules in a traditional VRLA bank already show internal sulfation that no external multimeter detects. When your vendor measures them “on-site”, the result depends on the voltage at that moment, not the degradation trajectory. Predictive maintenance with thermography + AI lets you see that trajectory before the bank fails during a real event.

Why replacing batteries “by age” no longer makes sense

The “replace every 3 years” rule comes from the manufacturer’s manual, not from the real state of your bank. In practice, VRLA batteries in data centers operate between 18°C and 27°C; every 8°C above 25°C halves useful life. That means a bank on a site with marginal cooling at 32°C ages at twice the speed of the same bank at 24°C.

The result: two identical UPS, bought the same day, end up with completely different aging patterns. The one on the hot site fails at 2 years. The one on the cold site keeps going for 5 years. Replacing both “by age” means throwing money away on one and leaving the other exposed to an unnecessary failure.

What predictive maintenance measures (and what it does not)

A PdM program combines three layers of continuous telemetry plus an AI model that cross-references them:

Internal impedance per cell — measures the module's internal resistance, an early indicator of sulfation and dry-out. The trend across weeks matters more than any single reading.
Float voltage and current drift — detects when a cell starts accepting less or more than the bank average, a sign of imminent failure.
Temperature per terminal — thermal imaging every quarter catches loose connections before they arc. Loose terminals are the second cause of bank failure after sulfation.

What PdM does NOT do: it does not predict an electronic board failure or a short circuit inside the UPS itself. Those need a different layer (UPS diagnostics + event logs).

What changes when you add AI to PdM

Manufacturer remote monitoring services have been collecting telemetry from UPS in the field for years. The breakthrough in the last 36 months has been the AI layer that learns the specific pattern of each bank and identifies drift that humans miss.

Eaton PredictPulse, for example, ingests data from over 100,000 UPS in the field and produces a health score per bank that anticipates replacement within ±60 days. Vertiv LIFE Services does the same for its installed base. Both replace the calendar-based replacement cycle with a model-based cycle.

The operational difference is concrete: under a classic scheme, your vendor measures every 6 months, replaces by age, and you still have 2 unplanned failures per year per 100 batteries. With AI PdM, the failure rate drops to 0.2 to 0.4 per year per 100, and replacements are planned within the maintenance window instead of during an incident.

How much does a real PdM cost (and save) in Mexico

A monthly PdM monitoring service with AI + quarterly thermography for a 40-battery VRLA bank on a Tier III data center costs between MXN $35,000 and MXN $75,000 per month (USD $2,000 to $4,300 at 17.5 MXN/USD, approximate, verify Banxico FIX on publication day). It replaces a quarterly preventive visit of MXN $18,000 to $30,000 plus the annual battery swap of MXN $240,000 to $480,000.

The ROI is calculated against three avoided costs:

Unplanned failure during an event: between MXN $120,000 and MXN $400,000 per hour of downtime for a mid-size DC, plus SLA penalties.
Damage to adjacent equipment: a battery that fails in short can take down the UPS, the PDU downstream, and in the worst case the load. Replacement of a UPS chassis runs MXN $800,000 to $2,500,000.
Emergency dispatch vs scheduled maintenance: emergency call is 3x to 5x more expensive than scheduled; thermography + AI lets you schedule 90% of replacements.

For a site where one hour of downtime costs over MXN $250,000 (USD $14,300 at 17.5 MXN/USD, approximate, verify Banxico FIX on publication day), one avoided incident per year covers the annual cost of the PdM service. The math turns positive from day one.