Introduction
Modular UPS architecture divides the power path into independent power, battery and control modules that can be hot-swapped while the system continues to supply the load. This design dramatically reduces Mean-Time-to-Repair (MTTR), but only if the operator can decide—within seconds—which battery module is unhealthy and why. The following text summarizes field-proven techniques that allow a technician to locate a defective battery module in less than five minutes without removing the wrong pack or shutting the bus down.

Typical Failure Signature of a Battery Module
A battery module in a modular UPS is normally a 20–60 Ah, 42–54 V lithium-ion or VRLA string with its own Battery Management System (BMS). The failures most frequently seen in the field are:
a) Internal cell open-circuit (voltage collapses under load)
b) Cell short-circuit (module voltage lower by 2 V or 3.6 V than nominal)
c) BMS Hall sensor drift (current reading offset >3 %)
d) MOSFET fuse-blown inside the module (zero charge current)
e) Over-temperature shutdown (heat run, fan blocked, >60 °C)
f) Capacity fade (<70 % of nameplate after 400 cycles)

Four-Layer Diagnostic Model
Layer-1: System-level alarms (UPS LCD / SNMP)
Layer-2: Module-level telemetry (voltage, current, temp, impedance)
Layer-3: Cell-level trending (SOC imbalance, ΔV >300 mV)
Layer-4: Waveform capture (millisecond resolution during fault)

Layer-1 – Use the UPS Alarm Register
Modern modular UPS (Eaton 9PX, Vertiv Liebert APM, APC Symmetra PX) publish a binary alarm register as part of their SNMP MIB or Modbus map.
Key OIDs to poll:

Layer-2 – Compare Module Telemetry
Open the UPS web GUI or use the vendor’s software (Eaton IPM, Vertiv LIFE, APC StruxureWare). Export the real-time table that contains, for every battery module:

Layer-3 – Deep Dive into Cell Imbalance
If the UPS allows cell-level access (most lithium-ion modules do), open the “cell voltage” page. A healthy module keeps the sixteen 3.6 V cells within 50 mV of each other during float.
Rule: max(cellV) − min(cellV) >300 mV → imbalance >8 % capacity.
If the imbalance is localised inside the previously flagged suspect-1 module, you have cross-verified the fault; if the imbalance is spread over two adjacent modules, the problem is more likely a loose interconnect than a single bad cell.

Layer-4 – Waveform Capture for Intermittent Faults
Some faults appear only during the millisecond transition from mains to battery. Use the built-in “fault recorder” function that is already present in many Chinese modular UPS platforms. The recorder continuously writes 500 µs samples to a ring RAM; when the DSP throws a fault code it freezes 200 ms of post-fault data. Download the COMTRADE file and look at:

Passive IR Scan – Optional but Fast
If the cabinet door can be opened safely, use a pocket thermal imager (FLIR ONE, ≤USD 300). Scan the battery drawers within 10 s. A module that is >6 °C hotter than its neighbours almost always contains a high-impedance cell or balancing MOSFET running continuously. Mark the hot drawer with tape; the IR image is admissible evidence for a warranty claim.

One-Minute “Swap-and-Watch” Test
When the above data still leave ambiguity (for example two modules show similar ΔV), execute a minimal-invasive test:

Note the instantaneous battery current on the UPS LCD.

Swap the positions of suspect-1 and its left neighbour (hot-swap, <30 s).

Watch the current redistributes: if the alarm follows the module, the module is bad; if the alarm stays in the slot, the slot wiring or back-plane is bad.

Automated Battery Self-Test – Final Confirmation
Trigger the “battery capacity test” from the front panel. A 20 % discharge is low-risk yet sufficient to expose a 30 % capacity fade. A module that drops its voltage below 42 V (for 48 V lithium) before the test ends is tagged “Replace”. Abort the test immediately if any cell goes below 2.5 V to avoid deep-discharge damage.

Common Field Mistakes to Avoid

Document and Close the Loop
After replacement, save the following in the CMMS (computerised maintenance management system):

Predictive Extension – Machine-Learning Overlay
Once 50 or more historical fault packages are available, train a gradient-boosting classifier using the features: ΔV, ΔT, impedance, cycle count, calendar age, ambient temperature. The model can forecast “probability of failure within 30 days” with ~87 % precision, allowing the site to order spare modules just-in-time and cut spare inventory by 40 %.

Summary Workflow ( Pocket Card )

Read alarm register → get slot X (30 s)

Export telemetry → flag outliers (60 s)

Check cell imbalance → confirm (60 s)

Optional IR scan → mark hot drawer (30 s)

Swap-and-watch → fault follows module? (60 s)

20 % self-test → final proof (5 min)